Doc. no. N1808=05-0068
Date: 2005-05-01
Project: Programming Language C++
Reply to: Matt Austern <austern@google.com>

C++ Standard Library Closed Issues List (Revision R36)

Reference ISO/IEC IS 14882:1998(E)

Also see:

This document contains only library issues which have been closed by the Library Working Group as duplicates or not defects. That is, issues which have a status of Dup or NAD. See the Library Active Issues List active issues and more information. See the Library Defect Reports List for issues considered defects. The introductory material in that document also applies to this document.

Revision History

Closed Issues


2. Auto_ptr conversions effects incorrect

Section: 20.4.5.3 [lib.auto.ptr.conv]  Status: NAD  Submitter: Nathan Myers  Date: 4 Dec 1997

Paragraph 1 in "Effects", says "Calls p->release()" where it clearly must be "Calls p.release()". (As it is, it seems to require using auto_ptr<>::operator-> to refer to X::release, assuming that exists.)

Proposed resolution:

Change 20.4.5.3 [lib.auto.ptr.conv] paragraph 1 Effects from "Calls p->release()" to "Calls p.release()".

Rationale:

Not a defect: the proposed change is already found in the standard. [Originally classified as a defect, later reclassified.]


4. Basic_string size_type and difference_type should be implementation defined

Section: 21.3 [lib.basic.string]  Status: NAD  Submitter: Beman Dawes  Date: 16 Nov 1997

In Morristown we changed the size_type and difference_type typedefs for all the other containers to implementation defined with a reference to 23.1 [lib.container.requirements]. This should probably also have been done for strings.

Proposed resolution:

Rationale:

Not a defect. [Originally classified as a defect, later reclassified.] basic_string, unlike the other standard library template containers, is severely constrained by its use of char_traits. Those types are dictated by the traits class, and are far from implementation defined.


6. File position not an offset unimplementable

Section: 27.4.3 [lib.fpos]  Status: NAD  Submitter: Matt Austern  Date: 15 Dec 1997

Table 88, in I/O, is too strict; it's unimplementable on systems where a file position isn't just an offset. It also never says just what fpos<> is really supposed to be. [Here's my summary, which Jerry agrees is more or less accurate. "I think I now know what the class really is, at this point: it's a magic cookie that encapsulates an mbstate_t and a file position (possibly represented as an fpos_t), it has syntactic support for pointer-like arithmetic, and implementors are required to have real, not just syntactic, support for arithmetic." This isn't standardese, of course.]

Proposed resolution:

Rationale:

Not a defect. The LWG believes that the Standard is already clear, and that the above summary is what the Standard in effect says.


10. Codecvt<>::do unclear

Section: 22.2.1.5.2 [lib.locale.codecvt.virtuals]  Status: Dup  Submitter: Matt Austern  Date: 14 Jan 1998

Section 22.2.1.5.2 says that codecvt<>::do_in and do_out should return the value noconv if "no conversion was needed". However, I don't see anything anywhere that defines what it means for a conversion to be needed or not needed. I can think of several circumstances where one might plausibly think that a conversion is not "needed", but I don't know which one is intended here.

Proposed resolution:

Rationale:

Duplicate. See issue 19.


12. Way objects hold allocators unclear

Section: 20.1.5 [lib.allocator.requirements]  Status: NAD  Submitter: Angelika Langer  Date: 23 Feb 1998

I couldn't find a statement in the standard saying whether the allocator object held by a container is held as a copy of the constructor argument or whether a pointer of reference is maintained internal. There is an according statement for compare objects and how they are maintained by the associative containers, but I couldn't find anything regarding allocators.

Did I overlook it? Is it an open issue or known defect? Or is it deliberately left unspecified?

Proposed resolution:

Rationale:

Not a defect. The LWG believes that the Standard is already clear.  See 23.1 [lib.container.requirements], paragraph 8.


43. Locale table correction

Section: 22.2.1.5.2 [lib.locale.codecvt.virtuals]  Status: Dup  Submitter: Brendan Kehoe  Date: 1 Jun 1998

Proposed resolution:

Rationale:

Duplicate. See issue 33.


45. Stringstreams read/write pointers initial position unclear

Section: 27.7.3 [lib.ostringstream]  Status: NAD  Submitter: Matthias Mueller  Date: 27 May 1998

In a comp.lang.c++.moderated Matthias Mueller wrote:

"We are not sure how to interpret the CD2 (see 27.2 [lib.iostream.forward], 27.7.3.1 [lib.ostringstream.cons], 27.7.1.1 [lib.stringbuf.cons]) with respect to the question as to what the correct initial positions of the write and  read pointers of a stringstream should be."

"Is it the same to output two strings or to initialize the stringstream with the first and to output the second?"

[PJ Plauger, Bjarne Stroustrup, Randy Smithey, Sean Corfield, and Jerry Schwarz have all offered opinions; see reflector messages lib-6518, 6519, 6520, 6521, 6523, 6524.]

Proposed resolution:

Rationale:

The LWG believes the Standard is correct as written. The behavior of stringstreams is consistent with fstreams, and there is a constructor which can be used to obtain the desired effect. This behavior is known to be different from strstreams.


58. Extracting a char from a wide-oriented stream

Section: 27.6.1.2.3 [lib.istream::extractors]  Status: NAD  Submitter: Matt Austern  Date: 1 Jul 1998

27.6.1.2.3 has member functions for extraction of signed char and unsigned char, both singly and as strings. However, it doesn't say what it means to extract a char from a basic_streambuf<charT, Traits>.

basic_streambuf, after all, has no members to extract a char, so basic_istream must somehow convert from charT to signed char or unsigned char. The standard doesn't say how it is to perform that conversion.

Proposed resolution:

Rationale:

The Standard is correct as written. There is no such extractor and this is the intent of the LWG.


65. Underspecification of strstreambuf::seekoff

Section: D.7.1.3 [depr.strstreambuf.virtuals]  Status: NAD  Submitter: Matt Austern  Date: 18 Aug 1998

The standard says how this member function affects the current stream position. (gptr or pptr) However, it does not say how this member function affects the beginning and end of the get/put area.

This is an issue when seekoff is used to position the get pointer beyond the end of the current read area. (Which is legal. This is implicit in the definition of seekhigh in D.7.1, paragraph 4.)

Proposed resolution:

Rationale:

The LWG agrees that seekoff() is underspecified, but does not wish to invest effort in this deprecated feature.


67. Setw useless for strings

Section: 21.3.7.9 [lib.string.io]  Status: Dup  Submitter: Steve Clamage  Date: 9 Jul 1998

In a comp.std.c++ posting Michel Michaud wrote: What should be output by:

   string text("Hello");
   cout << '[' << setw(10) << right << text << ']';

Shouldn't it be:

   [     Hello]

Another person replied: Actually, according to the FDIS, the width of the field should be the minimum of width and the length of the string, so the output shouldn't have any padding. I think that this is a typo, however, and that what is wanted is the maximum of the two. (As written, setw is useless for strings. If that had been the intent, one wouldn't expect them to have mentioned using its value.)

It's worth pointing out that this is a recent correction anyway; IIRC, earlier versions of the draft forgot to mention formatting parameters whatsoever.

Proposed resolution:

Rationale:

Duplicate. See issue 25.


72. Do_convert phantom member function

Section: 22.2.1.5 [lib.locale.codecvt]  Status: Dup  Submitter: Nathan Myers  Date: 24 Aug 1998

In 22.2.1.5 [lib.locale.codecvt] par 3, and in 22.2.1.5.2 [lib.locale.codecvt.virtuals] par 8, a nonexistent member function "do_convert" is mentioned. This member was replaced with "do_in" and "do_out", the proper referents in the contexts above.

Proposed resolution:

Rationale:

Duplicate: see issue 24.


73. is_open should be const

Section: 27.8.1 [lib.fstreams]  Status: NAD  Submitter: Matt Austern  Date: 27 Aug 1998

Classes basic_ifstream, basic_ofstream, and basic_fstream all have a member function is_open. It should be a const member function, since it does nothing but call one of basic_filebuf's const member functions.

Proposed resolution:

Rationale:

Not a defect. This is a deliberate feature; const streams would be meaningless.


77. Valarray operator[] const returning value

Section: 26.3.2.3 [lib.valarray.access]  Status: NAD Future  Submitter: Levente Farkas  Date: 9 Sep 1998

valarray:

    T operator[] (size_t) const;

why not

    const T& operator[] (size_t) const;

as in vector ???

One can't copy even from a const valarray eg:

    memcpy(ptr, &v[0], v.size() * sizeof(double));

[I] find this bug in valarray is very difficult.

Proposed resolution:

Rationale:

The LWG believes that the interface was deliberately designed that way. That is what valarray was designed to do; that's where the "value array" name comes from. LWG members further comment that "we don't want valarray to be a full STL container." 26.3.2.3 [lib.valarray.access] specifies properties that indicate "an absence of aliasing" for non-constant arrays; this allows optimizations, including special hardware optimizations, that are not otherwise possible.


81. Wrong declaration of slice operations

Section: 26.3.5 [lib.template.slice.array], 26.3.7 [lib.template.gslice.array], 26.3.8 [lib.template.mask.array], 26.3.9 [lib.template.indirect.array]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

Isn't the definition of copy constructor and assignment operators wrong?        Instead of

      slice_array(const slice_array&); 
      slice_array& operator=(const slice_array&);

IMHO they have to be

      slice_array(const slice_array<T>&); 
      slice_array& operator=(const slice_array<T>&);

Same for gslice_array.

Proposed resolution:

Rationale:

Not a defect. The Standard is correct as written.


82. Missing constant for set elements

Section: 23.1.2 [lib.associative.reqmts]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

Paragraph 5 specifies:

For set and multiset the value type is the same as the key type. For map and multimap it is equal to pair<const Key, T>.

Strictly speaking, this is not correct because for set and multiset the value type is the same as the constant key type.

Proposed resolution:

Rationale:

Not a defect. The Standard is correct as written; it uses a different mechanism (const &) for set and multiset. See issue 103 for a related issue.


84. Ambiguity with string::insert()

Section: 21.3.5 [lib.string.modifiers]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

If I try

    s.insert(0,1,' ');

  I get an nasty ambiguity. It might be

    s.insert((size_type)0,(size_type)1,(charT)' ');

which inserts 1 space character at position 0, or

    s.insert((char*)0,(size_type)1,(charT)' ')

which inserts 1 space character at iterator/address 0 (bingo!), or

    s.insert((char*)0, (InputIterator)1, (InputIterator)' ')

which normally inserts characters from iterator 1 to iterator ' '. But according to 23.1.1.9 (the "do the right thing" fix) it is equivalent to the second. However, it is still ambiguous, because of course I mean the first!

Proposed resolution:

Rationale:

Not a defect. The LWG believes this is a "genetic misfortune" inherent in the design of string and thus not a defect in the Standard as such .


85. String char types

Section: 21 [lib.strings]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

The standard seems not to require that charT is equivalent to traits::char_type. So, what happens if charT is not equivalent to traits::char_type?

Proposed resolution:

Rationale:

There is already wording in 21.1 [lib.char.traits] paragraph 3 that requires them to be the same.


87. Error in description of string::compare()

Section: 21.3.6.8 [lib.string::compare]  Status: Dup  Submitter: Nico Josuttis  Date: 29 Sep 1998

The following compare() description is obviously a bug:

int compare(size_type pos, size_type n1, 
            charT *s, size_type n2 = npos) const;

because without passing n2 it should compare up to the end of the string instead of comparing npos characters (which throws an exception)

Proposed resolution:

Rationale:

Duplicate; see issue 5.


88. Inconsistency between string::insert() and string::append()

Section: 21.3.5.4 [lib.string::insert], 21.3.5.2 [lib.string::append]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

Why does

  template<class InputIterator> 
       basic_string& append(InputIterator first, InputIterator last);

return a string, while

  template<class InputIterator> 
       void insert(iterator p, InputIterator first, InputIterator last);

returns nothing ?

Proposed resolution:

Rationale:

The LWG believes this stylistic inconsistency is not sufficiently serious to constitute a defect.


89. Missing throw specification for string::insert() and string::replace()

Section: 21.3.5.4 [lib.string::insert], 21.3.5.6 [lib.string::replace]  Status: Dup  Submitter: Nico Josuttis  Date: 29 Sep 1998

All insert() and replace() members for strings with an iterator as first argument lack a throw specification. The throw specification should probably be: length_error if size exceeds maximum.

Proposed resolution:

Rationale:

Considered a duplicate because it will be solved by the resolution of issue 83.


93. Incomplete Valarray Subset Definitions

Section: 26.3 [lib.numarray]  Status: NAD  Submitter: Nico Josuttis  Date: 29 Sep 1998

You can easily create subsets, but you can't easily combine them with other subsets. Unfortunately, you almost always needs an explicit type conversion to valarray. This is because the standard does not specify that valarray subsets provide the same operations as valarrays.

For example, to multiply two subsets and assign the result to a third subset, you can't write the following:

va[slice(0,4,3)] = va[slice(1,4,3)] * va[slice(2,4,3)];

Instead, you have to code as follows:

va[slice(0,4,3)] = static_cast<valarray<double> >(va[slice(1,4,3)]) * 
                   static_cast<valarray<double> >(va[slice(2,4,3)]);

This is tedious and error-prone. Even worse, it costs performance because each cast creates a temporary objects, which could be avoided without the cast.

Proposed resolution:

Extend all valarray subset types so that they offer all valarray operations.

Rationale:

This is not a defect in the Standard; it is a request for an extension.


94. May library implementors add template parameters to Standard Library classes?

Section: 17.4.4 [lib.conforming]  Status: NAD  Submitter: Matt Austern  Date: 22 Jan 1998

Is it a permitted extension for library implementors to add template parameters to standard library classes, provided that those extra parameters have defaults? For example, instead of defining template <class T, class Alloc = allocator<T> > class vector; defining it as template <class T, class Alloc = allocator<T>, int N = 1> class vector;

The standard may well already allow this (I can't think of any way that this extension could break a conforming program, considering that users are not permitted to forward-declare standard library components), but it ought to be explicitly permitted or forbidden.

comment from Steve Cleary via comp.std.c++:

I disagree [with the proposed resolution] for the following reason: consider user library code with template template parameters. For example, a user library object may be templated on the type of underlying sequence storage to use (deque/list/vector), since these classes all take the same number and type of template parameters; this would allow the user to determine the performance tradeoffs of the user library object. A similar example is a user library object templated on the type of underlying set storage (set/multiset) or map storage (map/multimap), which would allow users to change (within reason) the semantic meanings of operations on that object.

I think that additional template parameters should be forbidden in the Standard classes. Library writers don't lose any expressive power, and can still offer extensions because additional template parameters may be provided by a non-Standard implementation class:

 
   template <class T, class Allocator = allocator<T>, int N = 1>
   class __vector
   { ... };
   template <class T, class Allocator = allocator<T> >
   class vector: public __vector<T, Allocator>
   { ... };

Proposed resolution:

Add a new subclause [presumably 17.4.4.9] following 17.4.4.8 [lib.res.on.exception.handling]:

17.4.4.9 Template Parameters

A specialization of a template class described in the C++ Standard Library behaves the same as if the implementation declares no additional template parameters.

Footnote: Additional template parameters with default values are thus permitted.

Add "template parameters" to the list of subclauses at the end of 17.4.4 [lib.conforming] paragraph 1.

[Kona: The LWG agreed the standard needs clarification. After discussion with John Spicer, it seems added template parameters can be detected by a program using template-template parameters. A straw vote - "should implementors be allowed to add template parameters?" found no consensus ; 5 - yes, 7 - no.]

Rationale:

There is no ambiguity; the standard is clear as written. Library implementors are not permitted to add template parameters to standard library classes. This does not fall under the "as if" rule, so it would be permitted only if the standard gave explicit license for implementors to do this. This would require a change in the standard.

The LWG decided against making this change, because it would break user code involving template template parameters or specializations of standard library class templates.


95. Members added by the implementation

Section: 17.4.4.4 [lib.member.functions]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

In 17.3.4.4/2 vs 17.3.4.7/0 there is a hole; an implementation could add virtual members a base class and break user derived classes.

Example:

// implementation code:
struct _Base { // _Base is in the implementer namespace
        virtual void foo ();
};
class vector : _Base // deriving from a class is allowed
{ ... };

// user code:
class vector_checking : public vector 
{
        void foo (); // don't want to override _Base::foo () as the 
                     // user doesn't know about _Base::foo ()
};

Proposed resolution:

Clarify the wording to make the example illegal.

Rationale:

This is not a defect in the Standard.  The example is already illegal.  See 17.4.4.4 [lib.member.functions] paragraph 2.


97. Insert inconsistent definition

Section: 23 [lib.containers]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

insert(iterator, const value_type&) is defined both on sequences and on set, with unrelated semantics: insert here (in sequences), and insert with hint (in associative containers). They should have different names (B.S. says: do not abuse overloading).

Proposed resolution:

Rationale:

This is not a defect in the Standard. It is a genetic misfortune of the design, for better or for worse.


99. Reverse_iterator comparisons completely wrong

Section: 24.4.1.3.13 [lib.reverse.iter.op<]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

The <, >, <=, >= comparison operator are wrong: they return the opposite of what they should.

Note: same problem in CD2, these were not even defined in CD1. SGI STL code is correct; this problem is known since the Morristown meeting but there it was too late

Proposed resolution:

Rationale:

This is not a defect in the Standard. A careful reading shows the Standard is correct as written. A review of several implementations show that they implement exactly what the Standard says.


100. Insert iterators/ostream_iterators overconstrained

Section: 24.4.2 [lib.insert.iterators], 24.5.4 [lib.ostreambuf.iterator]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

Overspecified For an insert iterator it, the expression *it is required to return a reference to it. This is a simple possible implementation, but as the SGI STL documentation says, not the only one, and the user should not assume that this is the case.

Proposed resolution:

Rationale:

The LWG believes this causes no harm and is not a defect in the standard. The only example anyone could come up with caused some incorrect code to work, rather than the other way around.


101. No way to free storage for vector and deque

Section: 23.2.4 [lib.vector], 23.2.1 [lib.deque]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

Reserve can not free storage, unlike string::reserve

Proposed resolution:

Rationale:

This is not a defect in the Standard. The LWG has considered this issue in the past and sees no need to change the Standard. Deque has no reserve() member function. For vector, shrink-to-fit can be expressed in a single line of code (where v is vector<T>):

vector<T>(v).swap(v);  // shrink-to-fit v


102. Bug in insert range in associative containers

Section: 23.1.2 [lib.associative.reqmts]  Status: Dup  Submitter: AFNOR  Date: 7 Oct 1998

Table 69 of Containers say that a.insert(i,j) is linear if [i, j) is ordered. It seems impossible to implement, as it means that if [i, j) = [x], insert in an associative container is O(1)!

Proposed resolution:

N+log (size()) if [i,j) is sorted according to value_comp()

Rationale:

Subsumed by issue 264.


104. Description of basic_string::operator[] is unclear

Section: 21.3.4 [lib.string.access]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

It is not clear that undefined behavior applies when pos == size () for the non const version.

Proposed resolution:

Rewrite as: Otherwise, if pos > size () or pos == size () and the non-const version is used, then the behavior is undefined.

Rationale:

The Standard is correct. The proposed resolution already appears in the Standard.


105. fstream ctors argument types desired

Section: 27.8 [lib.file.streams]  Status: NAD Future  Submitter: AFNOR  Date: 7 Oct 1998

fstream ctors take a const char* instead of string.
fstream ctors can't take wchar_t

An extension to add a const wchar_t* to fstream would make the implementation non conforming.

Proposed resolution:

Rationale:

This is not a defect in the Standard. It might be an interesting extension for the next Standard.


107. Valarray constructor is strange

Section: 26.3.2 [lib.template.valarray]  Status: NAD  Submitter: AFNOR  Date: 7 Oct 1998

The order of the arguments is (elem, size) instead of the normal (size, elem) in the rest of the library. Since elem often has an integral or floating point type, both types are convertible to each other and reversing them leads to a well formed program.

Proposed resolution:

Inverting the arguments could silently break programs. Introduce the two signatures (const T&, size_t) and (size_t, const T&), but make the one we do not want private so errors result in a diagnosed access violation. This technique can also be applied to STL containers.

Rationale:

The LWG believes that while the order of arguments is unfortunate, it does not constitute a defect in the standard. The LWG believes that the proposed solution will not work for valarray<size_t> and perhaps other cases.


111. istreambuf_iterator::equal overspecified, inefficient

Section: 24.5.3.5 [lib.istreambuf.iterator::equal]  Status: NAD Future  Submitter: Nathan Myers  Date: 15 Oct 1998

The member istreambuf_iterator<>::equal is specified to be unnecessarily inefficient. While this does not affect the efficiency of conforming implementations of iostreams, because they can "reach into" the iterators and bypass this function, it does affect users who use istreambuf_iterators.

The inefficiency results from a too-scrupulous definition, which requires a "true" result if neither iterator is at eof. In practice these iterators can only usefully be compared with the "eof" value, so the extra test implied provides no benefit, but slows down users' code.

The solution is to weaken the requirement on the function to return true only if both iterators are at eof.

Proposed resolution:

Replace 24.5.3.5 [lib.istreambuf.iterator::equal], paragraph 1,

-1- Returns: true if and only if both iterators are at end-of-stream, or neither is at end-of-stream, regardless of what streambuf object they use.

with

-1- Returns: true if and only if both iterators are at end-of-stream, regardless of what streambuf object they use.

Rationale:

It is not clear that this is a genuine defect. Additionally, the LWG was reluctant to make a change that would result in operator== not being a equivalence relation. One consequence of this change is that an algorithm that's passed the range [i, i) would no longer treat it as an empty range.


113. Missing/extra iostream sync semantics

Section: 27.6.1.1 [lib.istream], 27.6.1.3 [lib.istream.unformatted]  Status: NAD  Submitter: Steve Clamage  Date: 13 Oct 1998

In 27.6.1.1, class basic_istream has a member function sync, described in 27.6.1.3, paragraph 36.

Following the chain of definitions, I find that the various sync functions have defined semantics for output streams, but no semantics for input streams. On the other hand, basic_ostream has no sync function.

The sync function should at minimum be added to basic_ostream, for internal consistency.

A larger question is whether sync should have assigned semantics for input streams.

Classic iostreams said streambuf::sync flushes pending output and attempts to return unread input characters to the source. It is a protected member function. The filebuf version (which is public) has that behavior (it backs up the read pointer). Class strstreambuf does not override streambuf::sync, and so sync can't be called on a strstream.

If we can add corresponding semantics to the various sync functions, we should. If not, we should remove sync from basic_istream.

Proposed resolution:

Rationale:

A sync function is not needed in basic_ostream because the flush function provides the desired functionality.

As for the other points, the LWG finds the standard correct as written.


116. bitset cannot be constructed with a const char*

Section: 23.3.5 [lib.template.bitset]  Status: NAD Future  Submitter: Judy Ward  Date: 6 Nov 1998

The following code does not compile with the EDG compiler:

#include <bitset>
using namespace std;
bitset<32> b("111111111");

If you cast the ctor argument to a string, i.e.:

bitset<32> b(string("111111111"));

then it will compile. The reason is that bitset has the following templatized constructor:

template <class charT, class traits, class Allocator>
explicit bitset (const basic_string<charT, traits, Allocator>& str, ...);

According to the compiler vendor, Steve Adamcyk at EDG, the user cannot pass this template constructor a const char* and expect a conversion to basic_string. The reason is "When you have a template constructor, it can get used in contexts where type deduction can be done. Type deduction basically comes up with exact matches, not ones involving conversions."

I don't think the intention when this constructor became templatized was for construction from a const char* to no longer work.

Proposed resolution:

Add to 23.3.5 [lib.template.bitset] a bitset constructor declaration

explicit bitset(const char*);

and in Section 23.3.5.1 [lib.bitset.cons] add:

explicit bitset(const char* str);

Effects:
    Calls bitset((string) str, 0, string::npos);

Rationale:

Although the problem is real, the standard is designed that way so it is not a defect. Education is the immediate workaround. A future standard may wish to consider the Proposed Resolution as an extension.


121. Detailed definition for ctype<wchar_t> specialization

Section: 22.1.1.1.1 [lib.locale.category]  Status: NAD  Submitter: Judy Ward  Date: 15 Dec 1998

Section 22.1.1.1.1 has the following listed in Table 51: ctype<char> , ctype<wchar_t>.

Also Section 22.2.1.1 [lib.locale.ctype] says:

The instantiations required in Table 51 (22.1.1.1.1) namely ctype<char> and ctype<wchar_t> , implement character classing appropriate to the implementation's native character set.

However, Section 22.2.1.3 [lib.facet.ctype.special] only has a detailed description of the ctype<char> specialization, not the ctype<wchar_t> specialization.

Proposed resolution:

Add the ctype<wchar_t> detailed class description to Section 22.2.1.3 [lib.facet.ctype.special].

Rationale:

Specialization for wchar_t is not needed since the default is acceptable.


128. Need open_mode() function for file stream, string streams, file buffers, and string  buffers

Section: 27.7 [lib.string.streams], 27.8 [lib.file.streams]  Status: NAD Future  Submitter: Angelika Langer  Date: 22 Feb 1999

The following question came from Thorsten Herlemann:

You can set a mode when constructing or opening a file-stream or filebuf, e.g. ios::in, ios::out, ios::binary, ... But how can I get that mode later on, e.g. in my own operator << or operator >> or when I want to check whether a file-stream or file-buffer object passed as parameter is opened for input or output or binary? Is there no possibility? Is this a design-error in the standard C++ library?

It is indeed impossible to find out what a stream's or stream buffer's open mode is, and without that knowledge you don't know how certain operations behave. Just think of the append mode.

Both streams and stream buffers should have a mode() function that returns the current open mode setting.

Proposed resolution:

For stream buffers, add a function to the base class as a non-virtual function qualified as const to 27.5.2 [lib.streambuf]:

    openmode mode() const;

    Returns the current open mode.

With streams, I'm not sure what to suggest. In principle, the mode could already be returned by ios_base, but the mode is only initialized for file and string stream objects, unless I'm overlooking anything. For this reason it should be added to the most derived stream classes. Alternatively, it could be added to basic_ios and would be default initialized in basic_ios<>::init().

Rationale:

This might be an interesting extension for some future, but it is not a defect in the current standard. The Proposed Resolution is retained for future reference.


131. list::splice throws nothing

Section: 23.2.2.4 [lib.list.ops]  Status: NAD  Submitter: Howard Hinnant  Date: 6 Mar 1999

What happens if a splice operation causes the size() of a list to grow beyond max_size()?

Proposed resolution:

Rationale:

Size() cannot grow beyond max_size(). 


135. basic_iostream doubly initialized

Section: 27.6.1.5.1 [lib.iostream.cons]  Status: NAD  Submitter: Howard Hinnant  Date: 6 Mar 1999

-1- Effects Constructs an object of class basic_iostream, assigning initial values to the base classes by calling basic_istream<charT,traits>(sb) (lib.istream) and basic_ostream<charT,traits>(sb) (lib.ostream)

The called for basic_istream and basic_ostream constructors call init(sb). This means that the basic_iostream's virtual base class is initialized twice.

Proposed resolution:

Change 27.6.1.5.1, paragraph 1 to:

-1- Effects Constructs an object of class basic_iostream, assigning initial values to the base classes by calling basic_istream<charT,traits>(sb) (lib.istream).

Rationale:

The LWG agreed that the init() function is called twice, but said that this is harmless and so not a defect in the standard.


138. Class ctype_byname<char> redundant and misleading

Section: 22.2.1.4 [lib.locale.ctype.byname.special]  Status: NAD Future  Submitter: Angelika Langer  Date: March 18, 1999

Section 22.2.1.4 [lib.locale.ctype.byname.special] specifies that ctype_byname<char> must be a specialization of the ctype_byname template.

It is common practice in the standard that specializations of class templates are only mentioned where the interface of the specialization deviates from the interface of the template that it is a specialization of. Otherwise, the fact whether or not a required instantiation is an actual instantiation or a specialization is left open as an implementation detail.

Clause 22.2.1.4 deviates from that practice and for that reason is misleading. The fact, that ctype_byname<char> is specified as a specialization suggests that there must be something "special" about it, but it has the exact same interface as the ctype_byname template. Clause 22.2.1.4 does not have any explanatory value, is at best redundant, at worst misleading - unless I am missing anything.

Naturally, an implementation will most likely implement ctype_byname<char> as a specialization, because the base class ctype<char> is a specialization with an interface different from the ctype template, but that's an implementation detail and need not be mentioned in the standard.

Proposed resolution:

Rationale:

The standard as written is mildly misleading, but the correct fix is to deal with the underlying problem in the ctype_byname base class, not in the specialization. See issue 228.


140. map<Key, T>::value_type does not satisfy the assignable requirement

Section: 23.3.1 [lib.map]  Status: NAD Future  Submitter: Mark Mitchell  Date: 14 Apr 1999

23.1 [lib.container.requirements]

expression         return type      pre/post-condition
-------------     -----------      -------------------
X::value_type    T                    T is assignable

23.3.1 [lib.map]

A map satisfies all the requirements of a container.

For a map<Key, T> ... the value_type is pair<const Key, T>.

There's a contradiction here. In particular, `pair<const Key, T>' is not assignable; the `const Key' cannot be assigned to. So,  map<Key, T>::value_type does not satisfy the assignable requirement imposed by a container.

[See issue 103 for the slightly related issue of modification of set keys.]

Proposed resolution:

Rationale:

The LWG believes that the standard is inconsistent, but that this is a design problem rather than a strict defect. May wish to reconsider for the next standard.


143. C .h header wording unclear

Section: D.5 [depr.c.headers]  Status: NAD  Submitter: Christophe de Dinechin  Date: 4 May 1999

[depr.c.headers] paragraph 2 reads:

Each C header, whose name has the form name.h, behaves as if each name placed in the Standard library namespace by the corresponding cname header is also placed within the namespace scope of the namespace std and is followed by an explicit using-declaration (_namespace.udecl_)

I think it should mention the global name space somewhere...  Currently, it indicates that name placed in std is also placed in std...

I don't know what is the correct wording. For instance, if struct tm is defined in time.h, ctime declares std::tm. However, the current wording seems ambiguous regarding which of the following would occur for use of both ctime and time.h:

// version 1:
namespace std {
        struct tm { ... };
}
using std::tm;

// version 2:
struct tm { ... };
namespace std {
        using ::tm;
}

// version 3:
struct tm { ... };
namespace std {
        struct tm { ... };
}

I think version 1 is intended.

[Kona: The LWG agreed that the wording is not clear. It also agreed that version 1 is intended, version 2 is not equivalent to version 1, and version 3 is clearly not intended. The example below was constructed by Nathan Myers to illustrate why version 2 is not equivalent to version 1.

Although not equivalent, the LWG is unsure if (2) is enough of a problem to be prohibited. Points discussed in favor of allowing (2):

]

Example:

#include <time.h>
#include <utility>

int main() {
    std::tm * t;
    make_pair( t, t ); // okay with version 1 due to Koenig lookup
                       // fails with version 2; make_pair not found
    return 0;
}

Proposed resolution:

Replace D.5 [depr.c.headers] paragraph 2 with:

Each C header, whose name has the form name.h, behaves as if each name placed in the Standard library namespace by the corresponding cname header is also placed within the namespace scope of the namespace std by name.h and is followed by an explicit using-declaration (_namespace.udecl_) in global scope.

Rationale:

The current wording in the standard is the result of a difficult compromise that averted delay of the standard. Based on discussions in Tokyo it is clear that there is no still no consensus on stricter wording, so the issue has been closed. It is suggested that users not write code that depends on Koenig lookup of C library functions.


145. adjustfield lacks default value

Section: 27.4.4.1 [lib.basic.ios.cons]  Status: NAD  Submitter: Angelika Langer  Date: 12 May 1999

There is no initial value for the adjustfield defined, although many people believe that the default adjustment were right. This is a common misunderstanding. The standard only defines that, if no adjustment is specified, all the predefined inserters must add fill characters before the actual value, which is "as if" the right flag were set. The flag itself need not be set.

When you implement a user-defined inserter you cannot rely on right being the default setting for the adjustfield. Instead, you must be prepared to find none of the flags set and must keep in mind that in this case you should make your inserter behave "as if" the right flag were set. This is surprising to many people and complicates matters more than necessary.

Unless there is a good reason why the adjustfield should not be initialized I would suggest to give it the default value that everybody expects anyway.

Proposed resolution:

Rationale:

This is not a defect. It is deliberate that the default is no bits set. Consider Arabic or Hebrew, for example. See 22.2.2.2.2 [lib.facet.num.put.virtuals] paragraph 19, Table 61 - Fill padding.


149. Insert should return iterator to first element inserted

Section: 23.1.1 [lib.sequence.reqmts]  Status: NAD Future  Submitter: Andrew Koenig  Date: 28 Jun 1999

Suppose that c and c1 are sequential containers and i is an iterator that refers to an element of c. Then I can insert a copy of c1's elements into c ahead of element i by executing

c.insert(i, c1.begin(), c1.end());

If c is a vector, it is fairly easy for me to find out where the newly inserted elements are, even though i is now invalid:

size_t i_loc = i - c.begin();
c.insert(i, c1.begin(), c1.end());

and now the first inserted element is at c.begin()+i_loc and one past the last is at c.begin()+i_loc+c1.size().

But what if c is a list? I can still find the location of one past the last inserted element, because i is still valid. To find the location of the first inserted element, though, I must execute something like

for (size_t n = c1.size(); n; --n)
   --i;

because i is now no longer a random-access iterator.

Alternatively, I might write something like

bool first = i == c.begin();
list<T>::iterator j = i;
if (!first) --j;
c.insert(i, c1.begin(), c1.end());
if (first)
   j = c.begin();
else
   ++j;

which, although wretched, requires less overhead.

But I think the right solution is to change the definition of insert so that instead of returning void, it returns an iterator that refers to the first element inserted, if any, and otherwise is a copy of its first argument. 

Proposed resolution:

Rationale:

The LWG believes this was an intentional design decision and so is not a defect. It may be worth revisiting for the next standard.


157. Meaningless error handling for pword() and iword()

Section: 27.4.2.5 [lib.ios.base.storage]  Status: Dup  Submitter: Dietmar Kühl  Date: 20 Jul 1999

According to paragraphs 2 and 4 of 27.4.2.5 [lib.ios.base.storage], the functions iword() and pword() "set the badbit (which might throw an exception)" on failure. ... but what does it mean for ios_base to set the badbit? The state facilities of the IOStream library are defined in basic_ios, a derived class! It would be possible to attempt a down cast but then it would be necessary to know the character type used...

Proposed resolution:

Rationale:

Duplicate. See issue 41.


162. Really "formatted input functions"?

Section: 27.6.1.2.3 [lib.istream::extractors]  Status: Dup  Submitter: Dietmar Kühl  Date: 20 Jul 1999

It appears to be somewhat nonsensical to consider the functions defined in the paragraphs 1 to 5 to be "Formatted input function" but since these functions are defined in a section labeled "Formatted input functions" it is unclear to me whether these operators are considered formatted input functions which have to conform to the "common requirements" from 27.6.1.2.1 [lib.istream.formatted.reqmts]: If this is the case, all manipulators, not just ws, would skip whitespace unless noskipws is set (... but setting noskipws using the manipulator syntax would also skip whitespace :-)

See also issue 166 for the same problem in formatted output

Proposed resolution:

Rationale:

Duplicate. See issue 60.


163. Return of gcount() after a call to gcount

Section: 27.6.1.3 [lib.istream.unformatted]  Status: Dup  Submitter: Dietmar Kühl  Date: 20 Jul 1999

It is not clear which functions are to be considered unformatted input functions. As written, it seems that all functions in 27.6.1.3 [lib.istream.unformatted] are unformatted input functions. However, it does not really make much sense to construct a sentry object for gcount(), sync(), ... Also it is unclear what happens to the gcount() if eg. gcount(), putback(), unget(), or sync() is called: These functions don't extract characters, some of them even "unextract" a character. Should this still be reflected in gcount()? Of course, it could be read as if after a call to gcount() gcount() return 0 (the last unformatted input function, gcount(), didn't extract any character) and after a call to putback() gcount() returns -1 (the last unformatted input function putback() did "extract" back into the stream). Correspondingly for unget(). Is this what is intended? If so, this should be clarified. Otherwise, a corresponding clarification should be used.

Proposed resolution:

Rationale:

Duplicate.  See issue 60.


166. Really "formatted output functions"?

Section: 27.6.2.5.3 [lib.ostream.inserters]  Status: Dup  Submitter: Dietmar Kühl  Date: 20 Jul 1999

From 27.6.2.5.1 [lib.ostream.formatted.reqmts] it appears that all the functions defined in 27.6.2.5.3 [lib.ostream.inserters] have to construct a sentry object. Is this really intended?

This is basically the same problem as issue 162 but for output instead of input.

Proposed resolution:

Rationale:

Duplicate. See issue 60.


177. Complex operators cannot be explicitly instantiated

Section: 26.2.6 [lib.complex.ops]  Status: NAD  Submitter: Judy Ward  Date: 2 Jul 1999

A user who tries to explicitly instantiate a complex non-member operator will get compilation errors. Below is a simplified example of the reason why. The problem is that iterator_traits cannot be instantiated on a non-pointer type like float, yet when the compiler is trying to decide which operator+ needs to be instantiated it must instantiate the declaration to figure out the first argument type of a reverse_iterator operator.

namespace std {
template <class Iterator> 
struct iterator_traits
{
    typedef typename Iterator::value_type value_type;
};

template <class T> class reverse_iterator;

// reverse_iterator operator+
template <class T> 
reverse_iterator<T> operator+
(typename iterator_traits<T>::difference_type, const reverse_iterator<T>&);

template <class T> struct complex {};

// complex operator +
template <class T>
complex<T> operator+ (const T& lhs, const complex<T>& rhs) 
{ return complex<T>();} 
}

// request for explicit instantiation
template std::complex<float> std::operator+<float>(const float&, 
     const std::complex<float>&);

See also c++-stdlib reflector messages: lib-6814, 6815, 6816.

Proposed resolution:

Rationale:

Implementors can make minor changes and the example will work. Users are not affected in any case.

According to John Spicer, It is possible to explicitly instantiate these operators using different syntax: change "std::operator+<float>" to "std::operator+".

The proposed resolution of issue 120 is that users will not be able to explicitly instantiate standard library templates. If that resolution is accepted then library implementors will be the only ones that will be affected by this problem, and they must use the indicated syntax.


178. Should clog and cerr initially be tied to cout?

Section: 27.3.1 [lib.narrow.stream.objects]  Status: NAD  Submitter: Judy Ward  Date: 2 Jul 1999

Section 27.3.1 says "After the object cerr is initialized, cerr.flags() & unitbuf is nonzero. Its state is otherwise the same as required for ios_base::init (lib.basic.ios.cons). It doesn't say anything about the the state of clog. So this means that calling cerr.tie() and clog.tie() should return 0 (see Table 89 for ios_base::init effects).

Neither of the popular standard library implementations that I tried does this, they both tie cerr and clog to &cout. I would think that would be what users expect.

Proposed resolution:

Rationale:

The standard is clear as written.

27.3.1/5 says that "After the object cerr is initialized, cerr.flags() & unitbuf is nonzero. Its state is otherwise the same as required for ios_base::init (27.4.4.1)." Table 89 in 27.4.4.1, which gives the postconditions of basic_ios::init(), says that tie() is 0. (Other issues correct ios_base::init to basic_ios::init().)


180. Container member iterator arguments constness has unintended consequences

Section: 23 [lib.containers]  Status: NAD Future  Submitter: Dave Abrahams  Date: 1 Jul 1999

It is the constness of the container which should control whether it can be modified through a member function such as erase(), not the constness of the iterators. The iterators only serve to give positioning information.

Here's a simple and typical example problem which is currently very difficult or impossible to solve without the change proposed below.

Wrap a standard container C in a class W which allows clients to find and read (but not modify) a subrange of (C.begin(), C.end()]. The only modification clients are allowed to make to elements in this subrange is to erase them from C through the use of a member function of W.

Proposed resolution:

Change all non-const iterator parameters of standard library container member functions to accept const_iterator parameters. Note that this change applies to all library clauses, including strings.

For example, in 21.3.5.5 change:

       iterator erase(iterator p);

to:
       iterator erase(const_iterator p);

Rationale:

The issue was discussed at length. It was generally agreed that 1) There is no major technical argument against the change (although there is a minor argument that some obscure programs may break), and 2) Such a change would not break const correctness. The concerns about making the change were 1) it is user detectable (although only in boundary cases), 2) it changes a large number of signatures, and 3) it seems more of a design issue that an out-and-out defect.

The LWG believes that this issue should be considered as part of a general review of const issues for the next revision of the standard. Also see issue 200.


188. valarray helpers missing augmented assignment operators

Section: 26.3.2.6 [lib.valarray.cassign]  Status: NAD Future  Submitter: Gabriel Dos Reis  Date: 15 Aug 1999

26.3.2.6 defines augmented assignment operators valarray<T>::op=(const T&), but fails to provide corresponding versions for the helper classes. Thus making the following illegal:

#include <valarray>

int main()
{
std::valarray<double> v(3.14, 1999);

v[99] *= 2.0; // Ok

std::slice s(0, 50, 2);

v[s] *= 2.0; // ERROR
}

I can't understand the intent of that omission. It makes the valarray library less intuitive and less useful.

Proposed resolution:

Rationale:

Although perhaps an unfortunate design decision, the omission is not a defect in the current standard.  A future standard may wish to add the missing operators.


190. min() and max() functions should be std::binary_functions

Section: 25.3.7 [lib.alg.min.max]  Status: NAD Future  Submitter: Mark Rintoul  Date: 26 Aug 1999

Both std::min and std::max are defined as template functions. This is very different than the definition of std::plus (and similar structs) which are defined as function objects which inherit std::binary_function.

This lack of inheritance leaves std::min and std::max somewhat useless in standard library algorithms which require a function object that inherits std::binary_function.

Proposed resolution:

Rationale:

Although perhaps an unfortunate design decision, the omission is not a defect in the current standard.  A future standard may wish to consider additional function objects.


191. Unclear complexity for algorithms such as binary search

Section: 25.3.3 [lib.alg.binary.search]  Status: NAD  Submitter: Nico Josuttis  Date: 10 Oct 1999

The complexity of binary_search() is stated as "At most log(last-first) + 2 comparisons", which seems to say that the algorithm has logarithmic complexity. However, this algorithms is defined for forward iterators. And for forward iterators, the need to step element-by-element results into linear complexity. But such a statement is missing in the standard. The same applies to lower_bound(), upper_bound(), and equal_range(). 

However, strictly speaking the standard contains no bug here. So this might considered to be a clarification or improvement.

Proposed resolution:

Rationale:

The complexity is expressed in terms of comparisons, and that complexity can be met even if the number of iterators accessed is linear. Paragraph 1 already says exactly what happens to iterators.


192. a.insert(p,t) is inefficient and overconstrained

Section: 23.1.2 [lib.associative.reqmts]  Status: NAD  Submitter: Ed Brey  Date: 6 Jun 1999

As defined in 23.1.2, paragraph 7 (table 69), a.insert(p,t) suffers from several problems:

expression return type pre/post-condition complexity
a.insert(p,t) iterator inserts t if and only if there is no element with key equivalent to the key of t in containers with unique keys; always inserts t in containers with equivalent keys. always returns the iterator pointing to the element with key equivalent to the key of t . iterator p is a hint pointing to where the insert should start to search. logarithmic in general, but amortized constant if t is inserted right after p .

1. For a container with unique keys, only logarithmic complexity is guaranteed if no element is inserted, even though constant complexity is always possible if p points to an element equivalent to t.

2. For a container with equivalent keys, the amortized constant complexity guarantee is only useful if no key equivalent to t exists in the container. Otherwise, the insertion could occur in one of multiple locations, at least one of which would not be right after p.

3. By guaranteeing amortized constant complexity only when t is inserted after p, it is impossible to guarantee constant complexity if t is inserted at the beginning of the container. Such a problem would not exist if amortized constant complexity was guaranteed if t is inserted before p, since there is always some p immediately before which an insert can take place.

4. For a container with equivalent keys, p does not allow specification of where to insert the element, but rather only acts as a hint for improving performance. This negates the added functionality that p would provide if it specified where within a sequence of equivalent keys the insertion should occur. Specifying the insert location provides more control to the user, while providing no disadvantage to the container implementation.

Proposed resolution:

In 23.1.2 [lib.associative.reqmts] paragraph 7, replace the row in table 69 for a.insert(p,t) with the following two rows:

expression return type pre/post-condition complexity
a_uniq.insert(p,t) iterator inserts t if and only if there is no element with key equivalent to the key of t. returns the iterator pointing to the element with key equivalent to the key of t. logarithmic in general, but amortized constant if t is inserted right before p or p points to an element with key equivalent to t.
a_eq.insert(p,t) iterator inserts t and returns the iterator pointing to the newly inserted element. t is inserted right before p if doing so preserves the container ordering. logarithmic in general, but amortized constant if t is inserted right before p.

Rationale:

Too big a change.  Furthermore, implementors report checking both before p and after p, and don't want to change this behavior.


194. rdbuf() functions poorly specified

Section: 27.4.4 [lib.ios]  Status: NAD  Submitter: Steve Clamage  Date: 7 Sep 1999

In classic iostreams, base class ios had an rdbuf function that returned a pointer to the associated streambuf. Each derived class had its own rdbuf function that returned a pointer of a type reflecting the actual type derived from streambuf. Because in ARM C++, virtual function overrides had to have the same return type, rdbuf could not be virtual.

In standard iostreams, we retain the non-virtual rdbuf function design, and in addition have an overloaded rdbuf function that sets the buffer pointer. There is no need for the second function to be virtual nor to be implemented in derived classes.

Minor question: Was there a specific reason not to make the original rdbuf function virtual?

Major problem: Friendly compilers warn about functions in derived classes that hide base-class overloads. Any standard implementation of iostreams will result in such a warning on each of the iostream classes, because of the ill-considered decision to overload rdbuf only in a base class.

In addition, users of the second rdbuf function must use explicit qualification or a cast to call it from derived classes. An explicit qualification or cast to basic_ios would prevent access to any later overriding version if there was one.

What I'd like to do in an implementation is add a using- declaration for the second rdbuf function in each derived class. It would eliminate warnings about hiding functions, and would enable access without using explicit qualification. Such a change I don't think would change the behavior of any valid program, but would allow invalid programs to compile:

 filebuf mybuf;
 fstream f;
 f.rdbuf(mybuf); // should be an error, no visible rdbuf

I'd like to suggest this problem as a defect, with the proposed resolution to require the equivalent of a using-declaration for the rdbuf function that is not replaced in a later derived class. We could discuss whether replacing the function should be allowed.

Proposed resolution:

Rationale:

For historical reasons, the standard is correct as written. There is a subtle difference between the base class rdbuf() and derived class rdbuf(). The derived class rdbuf() always returns the original streambuf, whereas the base class rdbuf() will return the "current streambuf" if that has been changed by the variant you mention.

Permission is not required to add such an extension. See 17.4.4.4 [lib.member.functions].


196. Placement new example has alignment problems

Section: 18.4.1.3 [lib.new.delete.placement]  Status: Dup  Submitter: Herb Sutter  Date: 15 Dec 1998

The example in 18.4.1.3 [lib.new.delete.placement] paragraph 4 reads:

[Example: This can be useful for constructing an object at a known address:

   char place[sizeof(Something)];
   Something* p = new (place) Something();

end example]

This example has potential alignment problems.

Proposed resolution:

Rationale:

Duplicate: see issue 114.


197. max_size() underspecified

Section: 20.1.5 [lib.allocator.requirements], 23.1 [lib.container.requirements]  Status: NAD  Submitter: Andy Sawyer  Date: 21 Oct 1999

Must the value returned by max_size() be unchanged from call to call?

Must the value returned from max_size() be meaningful?

Possible meanings identified in lib-6827:

1) The largest container the implementation can support given "best case" conditions - i.e. assume the run-time platform is "configured to the max", and no overhead from the program itself. This may possibly be determined at the point the library is written, but certainly no later than compile time.

2) The largest container the program could create, given "best case" conditions - i.e. same platform assumptions as (1), but take into account any overhead for executing the program itself. (or, roughly "storage=storage-sizeof(program)"). This does NOT include any resource allocated by the program. This may (or may not) be determinable at compile time.

3) The largest container the current execution of the program could create, given knowledge of the actual run-time platform, but again, not taking into account any currently allocated resource. This is probably best determined at program start-up.

4) The largest container the current execution program could create at the point max_size() is called (or more correctly at the point max_size() returns :-), given it's current environment (i.e. taking into account the actual currently available resources). This, obviously, has to be determined dynamically each time max_size() is called.

Proposed resolution:

Rationale:

max_size() isn't useful for very many things, and the existing wording is sufficiently clear for the few cases that max_size() can be used for. None of the attempts to change the existing wording were an improvement.

It is clear to the LWG that the value returned by max_size() can't change from call to call.


203. basic_istream::sentry::sentry() is uninstantiable with ctype<user-defined type>

Section: 27.6.1.1.2 [lib.istream::sentry]  Status: NAD  Submitter: Matt McClure and Dietmar Kühl  Date: 1 Jan 2000

27.6.1.1.2 Paragraph 4 states:

To decide if the character c is a whitespace character, the constructor performs ''as if'' it executes the following code fragment: 

const ctype<charT>& ctype = use_facet<ctype<charT> >(is.getloc());
if (ctype.is(ctype.space,c)!=0)
// c is a whitespace character.

But Table 51 in 22.1.1.1.1 only requires an implementation to provide specializations for ctype<char> and ctype<wchar_t>. If sentry's constructor is implemented using ctype, it will be uninstantiable for a user-defined character type charT, unless the implementation has provided non-working (since it would be impossible to define a correct ctype<charT> specialization for an arbitrary charT) definitions of ctype's virtual member functions.

It seems the intent the standard is that sentry should behave, in every respect, not just during execution, as if it were implemented using ctype, with the burden of providing a ctype specialization falling on the user. But as it is written, nothing requires the translation of sentry's constructor to behave as if it used the above code, and it would seem therefore, that sentry's constructor should be instantiable for all character types.

Note: If I have misinterpreted the intent of the standard with respect to sentry's constructor's instantiability, then a note should be added to the following effect:

An implementation is forbidden from using the above code if it renders the constructor uninstantiable for an otherwise valid character type.

In any event, some clarification is needed.

Proposed resolution:

Rationale:

It is possible but not easy to instantiate on types other than char or wchar_t; many things have to be done first. That is by intention and is not a defect.


204. distance(first, last) when "last" is before "first"

Section: 24.3.4 [lib.iterator.operations]  Status: NAD  Submitter: Rintala Matti  Date: 28 Jan 2000

Section 24.3.4 describes the function distance(first, last) (where first and last are iterators) which calculates "the number of increments or decrements needed to get from 'first' to 'last'".

The function should work for forward, bidirectional and random access iterators, and there is a requirement 24.3.4.5 which states that "'last' must be reachable from 'first'".

With random access iterators the function is easy to implement as "last - first".

With forward iterators it's clear that 'first' must point to a place before 'last', because otherwise 'last' would not be reachable from 'first'.

But what about bidirectional iterators? There 'last' is reachable from 'first' with the -- operator even if 'last' points to an earlier position than 'first'. However, I cannot see how the distance() function could be implemented if the implementation does not know which of the iterators points to an earlier position (you cannot use ++ or -- on either iterator if you don't know which direction is the "safe way to travel").

The paragraph 24.3.4.1 states that "for ... bidirectional iterators they use ++ to provide linear time implementations". However, the ++ operator is not mentioned in the reachability requirement. Furthermore 24.3.4.4 explicitly mentions that distance() returns the number of increments _or decrements_, suggesting that it could return a negative number also for bidirectional iterators when 'last' points to a position before 'first'.

Is a further requirement is needed to state that for forward and bidirectional iterators "'last' must be reachable from 'first' using the ++ operator". Maybe this requirement might also apply to random access iterators so that distance() would work the same way for every iterator category?

Proposed resolution:

Rationale:

"Reachable" is defined in the standard in 24.1 [lib.iterator.requirements] paragraph 6. The definition is only in terms of operator++(). The LWG sees no defect in the standard.


205.  numeric_limits unclear on how to determine floating point types

Section: 18.2.1.2 [lib.numeric.limits.members]  Status: NAD  Submitter: Steve Cleary  Date: 28 Jan 2000

In several places in 18.2.1.2 [lib.numeric.limits.members], a member is described as "Meaningful for all floating point types." However, no clear method of determining a floating point type is provided.

In 18.2.1.5 [lib.numeric.special], paragraph 1 states ". . . (for example, epsilon() is only meaningful if is_integer is false). . ." which suggests that a type is a floating point type if is_specialized is true and is_integer is false; however, this is unclear.

When clarifying this, please keep in mind this need of users: what exactly is the definition of floating point? Would a fixed point or rational representation be considered one? I guess my statement here is that there could also be types that are neither integer or (strictly) floating point.

Proposed resolution:

Rationale:

It is up to the implementor of a user define type to decide if it is a floating point type.


206. operator new(size_t, nothrow) may become unlinked to ordinary operator new if ordinary version replaced

Section: 18.4.1.1 [lib.new.delete.single]  Status: NAD  Submitter: Howard Hinnant  Date: 29 Aug 1999

As specified, the implementation of the nothrow version of operator new does not necessarily call the ordinary operator new, but may instead simply call the same underlying allocator and return a null pointer instead of throwing an exception in case of failure.

Such an implementation breaks code that replaces the ordinary version of new, but not the nothrow version. If the ordinary version of new/delete is replaced, and if the replaced delete is not compatible with pointers returned from the library versions of new, then when the replaced delete receives a pointer allocated by the library new(nothrow), crash follows.

The fix appears to be that the lib version of new(nothrow) must call the ordinary new. Thus when the ordinary new gets replaced, the lib version will call the replaced ordinary new and things will continue to work.

An alternative would be to have the ordinary new call new(nothrow). This seems sub-optimal to me as the ordinary version of new is the version most commonly replaced in practice. So one would still need to replace both ordinary and nothrow versions if one wanted to replace the ordinary version.

Another alternative is to put in clear text that if one version is replaced, then the other must also be replaced to maintain compatibility. Then the proposed resolution below would just be a quality of implementation issue. There is already such text in paragraph 7 (under the new(nothrow) version). But this nuance is easily missed if one reads only the paragraphs relating to the ordinary new.

Proposed resolution:

Rationale:

Yes, they may become unlinked, and that is by design. If a user replaces one, the user should also replace the other.


207. ctype<char> members return clause incomplete

Section: 22.2.1.3.2 [lib.facet.ctype.char.members]  Status: Dup  Submitter: Robert Klarer  Date: 2 Nov 1999

The widen and narrow member functions are described in 22.2.1.3.2, paragraphs 9-11. In each case we have two overloaded signatures followed by a Returns clause. The Returns clause only describes one of the overloads.

Proposed resolution:

Change the returns clause in 22.2.1.3.2 [lib.facet.ctype.char.members] paragraph 10 from:

    Returns: do_widen(low, high, to).

to:

    Returns: do_widen(c) or do_widen(low, high, to), respectively.

Change the returns clause in 22.2.1.3.2 [lib.facet.ctype.char.members] paragraph 11 from:

    Returns: do_narrow(low, high, to).

to:

    Returns: do_narrow(c) or do_narrow(low, high, to), respectively.

Rationale:

Subsumed by issue 153, which addresses the same paragraphs.


213. Math function overloads ambiguous

Section: 26.5 [lib.c.math]  Status: NAD  Submitter: Nico Josuttis  Date: 26 Feb 2000

Due to the additional overloaded versions of numeric functions for float and long double according to Section 26.5, calls such as int x; std::pow (x, 4) are ambiguous now in a standard conforming implementation. Current implementations solve this problem very different (overload for all types, don't overload for float and long double, use preprocessor, follow the standard and get ambiguities).

This behavior should be standardized or at least identified as implementation defined.

Proposed resolution:

Rationale:

These math issues are an understood and accepted consequence of the design. They have been discussed several times in the past. Users must write casts or write floating point expressions as arguments.


215. Can a map's key_type be const?

Section: 23.1.2 [lib.associative.reqmts]  Status: NAD  Submitter: Judy Ward  Date: 29 Feb 2000

A user noticed that this doesn't compile with the Rogue Wave library because the rb_tree class declares a key_allocator, and allocator<const int> is not legal, I think:

map < const int, ... > // legal?

which made me wonder whether it is legal for a map's key_type to be const. In email from Matt Austern he said:

I'm not sure whether it's legal to declare a map with a const key type. I hadn't thought about that question until a couple weeks ago. My intuitive feeling is that it ought not to be allowed, and that the standard ought to say so. It does turn out to work in SGI's library, though, and someone in the compiler group even used it. Perhaps this deserves to be written up as an issue too.

Proposed resolution:

Rationale:

The "key is assignable" requirement from table 69 in 23.1.2 [lib.associative.reqmts] already implies the key cannot be const.


216. setbase manipulator description flawed

Section: 27.6.3 [lib.std.manip]  Status: Dup  Submitter: Hyman Rosen  Date: 29 Feb 2000

27.6.3 [lib.std.manip] paragraph 5 says:

smanip setbase(int base);

Returns: An object s of unspecified type such that if out is an (instance of) basic_ostream then the expression out<<s behaves as if f(s) were called, in is an (instance of) basic_istream then the expression in>>s behaves as if f(s) were called. Where f can be defined as:

ios_base& f(ios_base& str, int base)
{
  // set basefield
  str.setf(n == 8 ? ios_base::oct :
                n == 10 ? ios_base::dec :
                n == 16 ? ios_base::hex :
                  ios_base::fmtflags(0), ios_base::basefield);
  return str;
}

There are two problems here. First, f takes two parameters, so the description needs to say that out<<s and in>>s behave as if f(s,base) had been called. Second, f is has a parameter named base, but is written as if the parameter was named n.

Actually, there's a third problem. The paragraph has grammatical errors. There needs to be an "and" after the first comma, and the "Where f" sentence fragment needs to be merged into its preceding sentence. You may also want to format the function a little better. The formatting above is more-or-less what the Standard contains.

Proposed resolution:

Rationale:

The resolution of this defect is subsumed by the proposed resolution for issue 193.

[Tokyo: The LWG agrees that this is a defect and notes that it occurs additional places in the section, all requiring fixes.]


218. Algorithms do not use binary predicate objects for default comparisons

Section: 25.3 [lib.alg.sorting]  Status: NAD  Submitter: Pablo Halpern  Date: 6 Mar 2000

Many of the algorithms take an argument, pred, of template parameter type BinaryPredicate or an argument comp of template parameter type Compare. These algorithms usually have an overloaded version that does not take the predicate argument. In these cases pred is usually replaced by the use of operator== and comp is replaced by the use of operator<.

This use of hard-coded operators is inconsistent with other parts of the library, particularly the containers library, where equality is established using equal_to<> and ordering is established using less<>. Worse, the use of operator<, would cause the following innocent-looking code to have undefined behavior:

vector<string*> vec;
sort(vec.begin(), vec.end());

The use of operator< is not defined for pointers to unrelated objects. If std::sort used less<> to compare elements, then the above code would be well-defined, since less<> is explicitly specialized to produce a total ordering of pointers.

Proposed resolution:

Rationale:

This use of operator== and operator< was a very deliberate, conscious, and explicitly made design decision; these operators are often more efficient. The predicate forms are available for users who don't want to rely on operator== and operator<.


219. find algorithm missing version that takes a binary predicate argument

Section: 25.1.2 [lib.alg.find]  Status: NAD Future  Submitter: Pablo Halpern  Date: 6 Mar 2000

The find function always searches for a value using operator== to compare the value argument to each element in the input iterator range. This is inconsistent with other find-related functions such as find_end and find_first_of, which allow the caller to specify a binary predicate object to be used for determining equality. The fact that this can be accomplished using a combination of find_if and bind_1st or bind_2nd does not negate the desirability of a consistent, simple, alternative interface to find.

Proposed resolution:

In section 25.1.2 [lib.alg.find], add a second prototype for find (between the existing prototype and the prototype for find_if), as follows:

    template<class InputIterator, class T, class BinaryPredicate>
      InputIterator find(InputIterator first, InputIterator last,
                         const T& value, BinaryPredicate bin_pred);

Change the description of the return from:

Returns: The first iterator i in the range [first, last) for which the following corresponding conditions hold: *i == value, pred(*i) != false. Returns last if no such iterator is found.

 to:

Returns: The first iterator i in the range [first, last) for which the following  corresponding condition holds: *i == value, bin_pred(*i,value) != false, pred(*) != false. Return last if no such iterator is found.

Rationale:

This is request for a pure extension, so it is not a defect in the current standard.  As the submitter pointed out, "this can be accomplished using a combination of find_if and bind_1st or bind_2nd".


236. ctype<char>::is() member modifies facet

Section: 22.2.1.3.2 [lib.facet.ctype.char.members]  Status: Dup  Submitter: Dietmar Kühl  Date: 24 Apr 2000

The description of the is() member in paragraph 4 of 22.2.1.3.2 [lib.facet.ctype.char.members] is broken: According to this description, the second form of the is() method modifies the masks in the ctype object. The correct semantics if, of course, to obtain an array of masks. The corresponding method in the general case, ie. the do_is() method as described in 22.2.1.1.2 [lib.locale.ctype.virtuals] paragraph 1 does the right thing.

Proposed resolution:

Change paragraph 4 from

The second form, for all *p in the range [low, high), assigns vec[p-low] to table()[(unsigned char)*p].

to become

The second form, for all *p in the range [low, high), assigns table()[(unsigned char)*p] to vec[p-low].

Rationale:

Duplicate. See issue 28.


244. Must find's third argument be CopyConstructible?

Section: 25.1.2 [lib.alg.find]  Status: NAD  Submitter: Andrew Koenig  Date: 02 May 2000

Is the following implementation of find acceptable?

        template<class Iter, class X>
        Iter find(Iter begin, Iter end, const X& x)
        {
            X x1 = x;           // this is the crucial statement
            while (begin != end && *begin != x1)
                ++begin;
            return begin;
        }

If the answer is yes, then it is implementation-dependent as to whether the following fragment is well formed:

        vector<string> v;

        find(v.begin(), v.end(), "foo");

At issue is whether there is a requirement that the third argument of find be CopyConstructible. There may be no problem here, but analysis is necessary.

Proposed resolution:

Rationale:

There is no indication in the standard that find's third argument is required to be Copy Constructible. The LWG believes that no such requirement was intended. As noted above, there are times when a user might reasonably pass an argument that is not Copy Constructible.


245. Which operations on istream_iterator trigger input operations?

Section: 24.5.1 [lib.istream.iterator]  Status: NAD  Submitter: Andrew Koenig  Date: 02 May 2000

I do not think the standard specifies what operation(s) on istream iterators trigger input operations. So, for example:

        istream_iterator<int> i(cin);

        int n = *i++;

I do not think it is specified how many integers have been read from cin. The number must be at least 1, of course, but can it be 2? More?

Proposed resolution:

Rationale:

The standard is clear as written: the stream is read every time operator++ is called, and it is also read either when the iterator is constructed or when operator* is called for the first time. In the example above, exactly two integers are read from cin.

There may be a problem with the interaction between istream_iterator and some STL algorithms, such as find. There are no guarantees about how many times find may invoke operator++.


246. a.insert(p,t) is incorrectly specified

Section: 23.1.2 [lib.associative.reqmts]  Status: Dup  Submitter: Mark Rodgers  Date: 19 May 2000

Closed issue 192 raised several problems with the specification of this function, but was rejected as Not A Defect because it was too big a change with unacceptable impacts on existing implementations. However, issues remain that could be addressed with a smaller change and with little or no consequent impact.

  1. The specification is inconsistent with the original proposal and with several implementations.

    The initial implementation by Hewlett Packard only ever looked immediately before p, and I do not believe there was any intention to standardize anything other than this behavior. Consequently, current implementations by several leading implementors also look immediately before p, and will only insert after p in logarithmic time. I am only aware of one implementation that does actually look after p, and it looks before p as well. It is therefore doubtful that existing code would be relying on the behavior defined in the standard, and it would seem that fixing this defect as proposed below would standardize existing practice.

  2. The specification is inconsistent with insertion for sequence containers.

    This is difficult and confusing to teach to newcomers. All insert operations that specify an iterator as an insertion location should have a consistent meaning for the location represented by that iterator.

  3. As specified, there is no way to hint that the insertion should occur at the beginning of the container, and the way to hint that it should occur at the end is long winded and unnatural.

    For a container containing n elements, there are n+1 possible insertion locations and n+1 valid iterators. For there to be a one-to-one mapping between iterators and insertion locations, the iterator must represent an insertion location immediately before the iterator.

  4. When appending sorted ranges using insert_iterators, insertions are guaranteed to be sub-optimal.

    In such a situation, the optimum location for insertion is always immediately after the element previously inserted. The mechanics of the insert iterator guarantee that it will try and insert after the element after that, which will never be correct. However, if the container first tried to insert before the hint, all insertions would be performed in amortized constant time.

Proposed resolution:

In 23.1.2 [lib.associative.reqmts] paragraph 7, table 69, make the following changes in the row for a.insert(p,t):

assertion/note pre/post condition:
Change the last sentence from

"iterator p is a hint pointing to where the insert should start to search."

to

"iterator p is a hint indicating that immediately before p may be a correct location where the insertion could occur."

complexity:
Change the words "right after" to "immediately before".

Rationale:

Duplicate; see issue 233.


249. Return Type of auto_ptr::operator=

Section: 20.4.5 [lib.auto.ptr]  Status: NAD  Submitter: Joseph Gottman  Date: 30 Jun 2000

According to section 20.4.5, the function auto_ptr::operator=() returns a reference to an auto_ptr. The reason that operator=() usually returns a reference is to facilitate code like

    int x,y,z;
    x = y = z = 1;

However, given analogous code for auto_ptrs,

    auto_ptr<int> x, y, z;
    z.reset(new int(1));
    x = y = z;

the result would be that z and y would both be set to NULL, instead of all the auto_ptrs being set to the same value. This makes such cascading assignments useless and counterintuitive for auto_ptrs.

Proposed resolution:

Change auto_ptr::operator=() to return void instead of an auto_ptr reference.

Rationale:

The return value has uses other than cascaded assignments: a user can call an auto_ptr member function, pass the auto_ptr to a function, etc. Removing the return value could break working user code.


255. Why do basic_streambuf<>::pbump() and gbump() take an int?

Section: 27.5.2 [lib.streambuf]  Status: NAD Future  Submitter: Martin Sebor  Date: 12 Aug 2000

The basic_streambuf members gbump() and pbump() are specified to take an int argument. This requirement prevents the functions from effectively manipulating buffers larger than std::numeric_limits<int>::max() characters. It also makes the common use case for these functions somewhat difficult as many compilers will issue a warning when an argument of type larger than int (such as ptrdiff_t on LLP64 architectures) is passed to either of the function. Since it's often the result of the subtraction of two pointers that is passed to the functions, a cast is necessary to silence such warnings. Finally, the usage of a native type in the functions signatures is inconsistent with other member functions (such as sgetn() and sputn()) that manipulate the underlying character buffer. Those functions take a streamsize argument.

Proposed resolution:

Change the signatures of these functions in the synopsis of template class basic_streambuf (27.5.2) and in their descriptions (27.5.2.3.1, p4 and 27.5.2.3.2, p4) to take a streamsize argument.

Although this change has the potential of changing the ABI of the library, the change will affect only platforms where int is different than the definition of streamsize. However, since both functions are typically inline (they are on all known implementations), even on such platforms the change will not affect any user code unless it explicitly relies on the existing type of the functions (e.g., by taking their address). Such a possibility is IMO quite remote.

Alternate Suggestion from Howard Hinnant, c++std-lib-7780:

This is something of a nit, but I'm wondering if streamoff wouldn't be a better choice than streamsize. The argument to pbump and gbump MUST be signed. But the standard has this to say about streamsize (27.4.1/2/Footnote):

[Footnote: streamsize is used in most places where ISO C would use size_t. Most of the uses of streamsize could use size_t, except for the strstreambuf constructors, which require negative values. It should probably be the signed type corresponding to size_t (which is what Posix.2 calls ssize_t). --- end footnote]

This seems a little weak for the argument to pbump and gbump. Should we ever really get rid of strstream, this footnote might go with it, along with the reason to make streamsize signed.

Rationale:

The LWG believes this change is too big for now. We may wish to reconsider this for a future revision of the standard. One possibility is overloading pbump, rather than changing the signature.


257. STL functional object and iterator inheritance.

Section: 20.3.1 [lib.base], 24.3.2 [lib.iterator.basic]  Status: NAD  Submitter: Robert Dick   Date: 17 Aug 2000

According to the November 1997 Draft Standard, the results of deleting an object of a derived class through a pointer to an object of its base class are undefined if the base class has a non-virtual destructor. Therefore, it is potentially dangerous to publicly inherit from such base classes.

Defect:
The STL design encourages users to publicly inherit from a number of classes which do nothing but specify interfaces, and which contain non-virtual destructors.

Attribution:
Wil Evers and William E. Kempf suggested this modification for functional objects.

Proposed resolution:

When a base class in the standard library is useful only as an interface specifier, i.e., when an object of the class will never be directly instantiated, specify that the class contains a protected destructor. This will prevent deletion through a pointer to the base class without performance, or space penalties (on any implementation I'm aware of).

As an example, replace...

    template <class Arg, class Result>
    struct unary_function {
            typedef Arg    argument_type;
            typedef Result result_type;
    };

... with...

    template <class Arg, class Result>
    struct unary_function {
            typedef Arg    argument_type;
            typedef Result result_type;
    protected:
            ~unary_function() {}
    };

Affected definitions:
 20.3.1 [lib.function.objects] -- unary_function, binary_function
 24.3.2 [lib.iterator.basic] -- iterator

Rationale:

The standard is clear as written; this is a request for change, not a defect in the strict sense. The LWG had several different objections to the proposed change. One is that it would prevent users from creating objects of type unary_function and binary_function. Doing so can sometimes be legitimate, if users want to pass temporaries as traits or tag types in generic code.


267. interaction of strstreambuf::overflow() and seekoff()

Section: D.7.1.3 [depr.strstreambuf.virtuals]  Status: NAD  Submitter: Martin Sebor  Date: 5 Oct 2000

It appears that the interaction of the strstreambuf members overflow() and seekoff() can lead to undefined behavior in cases where defined behavior could reasonably be expected. The following program demonstrates this behavior:

    #include <strstream>

    int main ()
    {
         std::strstreambuf sb;
         sb.sputc ('c');

         sb.pubseekoff (-1, std::ios::end, std::ios::in);
         return !('c' == sb.sgetc ());
    }

D.7.1.1, p1 initializes strstreambuf with a call to basic_streambuf<>(), which in turn sets all pointers to 0 in 27.5.2.1, p1.

27.5.2.2.5, p1 says that basic_streambuf<>::sputc(c) calls overflow(traits::to_int_type(c)) if a write position isn't available (it isn't due to the above).

D.7.1.3, p3 says that strstreambuf::overflow(off, ..., ios::in) makes at least one write position available (i.e., it allows the function to make any positive number of write positions available).

D.7.1.3, p13 computes newoff = seekhigh - eback(). In D.7.1, p4 we see seekhigh = epptr() ? epptr() : egptr(), or seekhigh = epptr() in this case. newoff is then epptr() - eback().

D.7.1.4, p14 sets gptr() so that gptr() == eback() + newoff + off, or gptr() == epptr() + off holds.

If strstreambuf::overflow() made exactly one write position available then gptr() will be set to just before epptr(), and the program will return 0. Buf if the function made more than one write position available, epptr() and gptr() will both point past pptr() and the behavior of the program is undefined.

Proposed resolution:

Change the last sentence of D.7.1 [depr.strstreambuf] paragraph 4 from

Otherwise, seeklow equals gbeg and seekhigh is either pend, if pend is not a null pointer, or gend.

to become

Otherwise, seeklow equals gbeg and seekhigh is either gend if 0 == pptr(), or pbase() + max where max is the maximum value of pptr() - pbase() ever reached for this stream.

[ pre-Copenhagen: Dietmar provided wording for proposed resolution. ]

[ post-Copenhagen: Fixed a typo: proposed resolution said to fix 4.7.1, not D.7.1. ]

Rationale:

This is related to issue 65: it's not clear what it means to seek beyond the current area. Without resolving issue 65 we can't resolve this. As with issue 65, the library working group does not wish to invest time nailing down corner cases in a deprecated feature.


269. cstdarg and unnamed parameters

Section: 18.7 [lib.support.runtime]  Status: NAD  Submitter: J. Stephen Adamczyk  Date: 10 Oct 2000

One of our customers asks whether this is valid C++:

   #include <cstdarg>

   void bar(const char *, va_list);

   void
   foo(const char *file, const char *, ...)
   {
     va_list ap;
     va_start(ap, file);
     bar(file, ap);
     va_end(ap);
   }

The issue being whether it is valid to use cstdarg when the final parameter before the "..." is unnamed. cstdarg is, as far as I can tell, inherited verbatim from the C standard. and the definition there (7.8.1.1 in the ISO C89 standard) refers to "the identifier of the rightmost parameter". What happens when there is no such identifier?

My personal opinion is that this should be allowed, but some tweak might be required in the C++ standard.

Proposed resolution:

Rationale:

Not a defect, the C and C++ standards are clear. It is impossible to use varargs if the parameter immediately before "..." has no name, because that is the parameter that must be passed to va_start. The example given above is broken, because va_start is being passed the wrong parameter.

There is no support for extending varargs to provide additional functionality beyond what's currently there. For reasons of C/C++ compatibility, it is especially important not to make gratuitous changes in this part of the C++ standard. The C committee has already been requested not to touch this part of the C standard unless necessary.


277. Normative encouragement in allocator requirements unclear

Section: 20.1.5 [lib.allocator.requirements]  Status: NAD  Submitter: Matt Austern  Date: 07 Nov 2000

In 20.1.5, paragraph 5, the standard says that "Implementors are encouraged to supply libraries that can accept allocators that encapsulate more general memory models and that support non-equal instances." This is intended as normative encouragement to standard library implementors. However, it is possible to interpret this sentence as applying to nonstandard third-party libraries.

Proposed resolution:

In 20.1.5, paragraph 5, change "Implementors" to "Implementors of the library described in this International Standard".

Rationale:

The LWG believes the normative encouragement is already sufficiently clear, and that there are no important consequences even if it is misunderstood.


279. const and non-const iterators should have equivalent typedefs

Section: 23.1 [lib.container.requirements]  Status: NAD  Submitter: Steve Cleary  Date: 27 Nov 2000

This came from an email from Steve Cleary to Fergus in reference to issue 179. The library working group briefly discussed this in Toronto and believes it should be a separate issue.

Steve said: "We may want to state that the const/non-const iterators must have the same difference type, size_type, and category."

(Comment from Judy) I'm not sure if the above sentence should be true for all const and non-const iterators in a particular container, or if it means the container's iterator can't be compared with the container's const_iterator unless the above it true. I suspect the former.

Proposed resolution:

In Section: 23.1 [lib.container.requirements], table 65, in the assertion/note pre/post condition for X::const_iterator, add the following:

typeid(X::const_iterator::difference_type) == typeid(X::iterator::difference_type)

typeid(X::const_iterator::size_type) == typeid(X::iterator::size_type)

typeid(X::const_iterator::category) == typeid(X::iterator::category)

Rationale:

Going through the types one by one: Iterators don't have a size_type. We already know that the difference types are identical, because the container requirements already say that the difference types of both X::iterator and X::const_iterator are both X::difference_type. The standard does not require that X::iterator and X::const_iterator have the same iterator category, but the LWG does not see this as a defect: it's possible to imagine cases in which it would be useful for the categories to be different.

It may be desirable to require X::iterator and X::const_iterator to have the same value type, but that is a new issue. (Issue 322.)


287. conflicting ios_base fmtflags

Section: 27.4.2.2 [lib.fmtflags.state]  Status: NAD  Submitter: Judy Ward  Date: 30 Dec 2000

The Effects clause for ios_base::setf(fmtflags fmtfl) says "Sets fmtfl in flags()". What happens if the user first calls ios_base::scientific and then calls ios_base::fixed or vice-versa? This is an issue for all of the conflicting flags, i.e. ios_base::left and ios_base::right or ios_base::dec, ios_base::hex and ios_base::oct.

I see three possible solutions:

  1. Set ios_base::failbit whenever the user specifies a conflicting flag with one previously explicitly set. If the constructor is supposed to set ios_base::dec (see discussion below), then the user setting hex or oct format after construction will not set failbit.
  2. The last call to setf "wins", i.e. it clears any conflicting previous setting.
  3. All the flags that the user specifies are set, but when actually interpreting them, fixed always override scientific, right always overrides left, dec overrides hex which overrides oct.

Most existing implementations that I tried seem to conform to resolution #3, except that when using the iomanip manipulator hex or oct then that always overrides dec, but calling setf(ios_base::hex) doesn't.

There is a sort of related issue, which is that although the ios_base constructor says that each ios_base member has an indeterminate value after construction, all the existing implementations I tried explicitly set ios_base::dec.

Proposed resolution:

Rationale:

adjustfield, basefield, and floatfield are each multi-bit fields. It is possible to set multiple bits within each of those fields. (For example, dec and oct). These fields are used by locale facets. The LWG reviewed the way in which each of those three fields is used, and believes that in each case the behavior is well defined for any possible combination of bits. See for example Table 58, in 22.2.2.2.2 [lib.facet.num.put.virtuals], noting the requirement in paragraph 6 of that section.

Users are advised to use manipulators, or else use the two-argument version of setf, to avoid unexpected behavior.


289. <cmath> requirements missing C float and long double versions

Section: 26.5 [lib.c.math]  Status: NAD  Submitter: Judy Ward  Date: 30 Dec 2000

In ISO/IEC 9899:1990 Programming Languages C we find the following concerning <math.h>:

7.13.4 Mathematics <math.h>
The names of all existing functions declared in the <math.h> header, suffixed with f or l, are reserved respectively for corresponding functions with float and long double arguments are return values.

For example, float sinf(float) is reserved.

In the C99 standard, <math.h> must contain declarations for these functions.

So, is it acceptable for an implementor to add these prototypes to the C++ versions of the math headers? Are they required?

Proposed resolution:

Add these Functions to Table 80, section 26.5 and to Table 99, section C.2:

    acosf asinf atanf atan2f ceilf cosf coshf 
    expf fabsf floorf fmodf frexpf ldexpf 
    logf log10f modff powf sinf sinhf sqrtf 
    tanf tanhf 
    acosl asinl atanl atan2l ceill cosl coshl 
    expl fabsl floorl fmodl frexpl ldexpl 
    logl log10l modfl powl sinl sinhl sqrtl 
    tanl tanhl

There should probably be a note saying that these functions are optional and, if supplied, should match the description in the 1999 version of the C standard. In the next round of C++ standardization they can then become mandatory.

Rationale:

The C90 standard, as amended, already permits (but does not require) these functions, and the C++ standard incorporates the C90 standard by reference. C99 is not an issue, because it is never referred to by the C++ standard.


293. Order of execution in transform algorithm

Section: 25.2.3 [lib.alg.transform]  Status: NAD  Submitter: Angelika Langer  Date: 04 Jan 2001

This issue is related to issue 242. In case that the resolution proposed for issue 242 is accepted, we have have the following situation: The 4 numeric algorithms (accumulate and consorts) as well as transform would allow a certain category of side effects. The numeric algorithms specify that they invoke the functor "for every iterator i in the range [first, last) in order". transform, in contrast, would not give any guarantee regarding order of invocation of the functor, which means that the functor can be invoked in any arbitrary order.

Why would that be a problem? Consider an example: say the transformator that is a simple enumerator ( or more generally speaking, "is order-sensitive" ). Since a standard compliant implementation of transform is free to invoke the enumerator in no definite order, the result could be a garbled enumeration. Strictly speaking this is not a problem, but it is certainly at odds with the prevalent understanding of transform as an algorithms that assigns "a new _corresponding_ value" to the output elements.

All implementations that I know of invoke the transformator in definite order, namely starting from first and proceeding to last - 1. Unless there is an optimization conceivable that takes advantage of the indefinite order I would suggest to specify the order, because it eliminate the uncertainty that users would otherwise have regarding the order of execution of their potentially order-sensitive function objects.

Proposed resolution:

In section 25.2.3 - Transform [lib.alg.transform] change:

-1- Effects: Assigns through every iterator i in the range [result, result + (last1 - first1)) a new corresponding value equal to op(*(first1 + (i - result)) or binary_op(*(first1 + (i - result), *(first2 + (i - result))).

to:

-1- Effects: Computes values by invoking the operation op or binary_op for every iterator in the range [first1, last1) in order. Assigns through every iterator i in the range [result, result + (last1 - first1)) a new corresponding value equal to op(*(first1 + (i - result)) or binary_op(*(first1 + (i - result), *(first2 + (i - result))).

Rationale:

For Input Iterators an order is already guaranteed, because only one order is possible. If a user who passes a Forward Iterator to one of these algorithms really needs a specific order of execution, it's possible to achieve that effect by wrapping it in an Input Iterator adaptor.


296. Missing descriptions and requirements of pair operators

Section: 20.2.2 [lib.pairs]  Status: NAD  Submitter: Martin Sebor  Date: 14 Jan 2001

The synopsis of the header <utility> in 20.2 [lib.utility] lists the complete set of equality and relational operators for pair but the section describing the template and the operators only describes operator==() and operator<(), and it fails to mention any requirements on the template arguments. The remaining operators are not mentioned at all.

Proposed resolution:

Rationale:

20.2.1 [lib.operators] paragraph 10 already specifies the semantics. That paragraph says that, if declarations of operator!=, operator>, operator<=, and operator>= appear without definitions, they are defined as specified in 20.2.1 [lib.operators]. There should be no user confusion, since that paragraph happens to immediately precede the specification of pair.


302. Need error indication from codecvt<>::do_length

Section: 22.2.1.5.2 [lib.locale.codecvt.virtuals]  Status: NAD  Submitter: Gregory Bumgardner  Date: 25 Jan 2001

The effects of codecvt<>::do_length() are described in 22.2.1.5.2, paragraph 10. As implied by that paragraph, and clarified in issue 75, codecvt<>::do_length() must process the source data and update the stateT argument just as if the data had been processed by codecvt<>::in(). However, the standard does not specify how do_length() would report a translation failure, should the source sequence contain untranslatable or illegal character sequences.

The other conversion methods return an "error" result value to indicate that an untranslatable character has been encountered, but do_length() already has a return value (the number of source characters that have been processed by the method).

Proposed resolution:

This issue cannot be resolved without modifying the interface. An exception cannot be used, as there would be no way to determine how many characters have been processed and the state object would be left in an indeterminate state.

A source compatible solution involves adding a fifth argument to length() and do_length() that could be used to return position of the offending character sequence. This argument would have a default value that would allow it to be ignored:

  int length(stateT& state, 
             const externT* from, 
             const externT* from_end, 
             size_t max,
             const externT** from_next = 0);

  virtual
  int do_length(stateT& state, 
                const externT* from, 
                const externT* from_end, 
                size_t max,
                const externT** from_next);

Then an exception could be used to report any translation errors and the from_next argument, if used, could then be used to retrieve the location of the offending character sequence.

Rationale:

The standard is already clear: the return value is the number of "valid complete characters". If it encounters an invalid sequence of external characters, it stops.


304. Must *a return an lvalue when a is an input iterator?

Section: 24.1 [lib.iterator.requirements]  Status: NAD  Submitter: Dave Abrahams  Date: 5 Feb 2001

We all "know" that input iterators are allowed to produce values when dereferenced of which there is no other in-memory copy.

But: Table 72, with a careful reading, seems to imply that this can only be the case if the value_type has no members (e.g. is a built-in type).

The problem occurs in the following entry:

  a->m     pre: (*a).m is well-defined
           Equivalent to (*a).m

*a.m can be well-defined if *a is not a reference type, but since operator->() must return a pointer for a->m to be well-formed, it needs something to return a pointer to. This seems to indicate that *a must be buffered somewhere to make a legal input iterator.

I don't think this was intentional.

Proposed resolution:

Rationale:

The current standard is clear and consistent. Input iterators that return rvalues are in fact implementable. They may in some cases require extra work, but it is still possible to define an operator-> in such cases: it doesn't have to return a T*, but may return a proxy type. No change to the standard is justified.


309. Does sentry catch exceptions?

Section: 27.6 [lib.iostream.format]  Status: NAD  Submitter: Martin Sebor  Date: 19 Mar 2001

The descriptions of the constructors of basic_istream<>::sentry (27.6.1.1.2 [lib.istream::sentry]) and basic_ostream<>::sentry (27.6.2.3 [lib.ostream::sentry]) do not explain what the functions do in case an exception is thrown while they execute. Some current implementations allow all exceptions to propagate, others catch them and set ios_base::badbit instead, still others catch some but let others propagate.

The text also mentions that the functions may call setstate(failbit) (without actually saying on what object, but presumably the stream argument is meant). That may have been fine for basic_istream<>::sentry prior to issue 195, since the function performs an input operation which may fail. However, issue 195 amends 27.6.1.1.2 [lib.istream::sentry], p2 to clarify that the function should actually call setstate(failbit | eofbit), so the sentence in p3 is redundant or even somewhat contradictory.

The same sentence that appears in 27.6.2.3 [lib.ostream::sentry], p3 doesn't seem to be very meaningful for basic_istream<>::sentry which performs no input. It is actually rather misleading since it would appear to guide library implementers to calling setstate(failbit) when os.tie()->flush(), the only called function, throws an exception (typically, it's badbit that's set in response to such an event).

Additional comments from Martin, who isn't comfortable with the current proposed resolution (see c++std-lib-11530)

The istream::sentry ctor says nothing about how the function deals with exemptions (27.6.1.1.2, p1 says that the class is responsible for doing "exception safe"(*) prefix and suffix operations but it doesn't explain what level of exception safety the class promises to provide). The mockup example of a "typical implementation of the sentry ctor" given in 27.6.1.1.2, p6, removed in ISO/IEC 14882:2003, doesn't show exception handling, either. Since the ctor is not classified as a formatted or unformatted input function, the text in 27.6.1.1, p1 through p4 does not apply. All this would seem to suggest that the sentry ctor should not catch or in any way handle exceptions thrown from any functions it may call. Thus, the typical implementation of an istream extractor may look something like [1].

The problem with [1] is that while it correctly sets ios::badbit if an exception is thrown from one of the functions called from the sentry ctor, if the sentry ctor reaches EOF while extracting whitespace from a stream that has eofbit or failbit set in exceptions(), it will cause an ios::failure to be thrown, which will in turn cause the extractor to set ios::badbit.

The only straightforward way to prevent this behavior is to move the definition of the sentry object in the extractor above the try block (as suggested by the example in 22.2.8, p9 and also indirectly supported by 27.6.1.3, p1). See [2]. But such an implementation will allow exceptions thrown from functions called from the ctor to freely propagate to the caller regardless of the setting of ios::badbit in the stream object's exceptions().

So since neither [1] nor [2] behaves as expected, the only possible solution is to have the sentry ctor catch exceptions thrown from called functions, set badbit, and propagate those exceptions if badbit is also set in exceptions(). (Another solution exists that deals with both kinds of sentries, but the code is non-obvious and cumbersome -- see [3].)

Please note that, as the issue points out, current libraries do not behave consistently, suggesting that implementors are not quite clear on the exception handling in istream::sentry, despite the fact that some LWG members might feel otherwise. (As documented by the parenthetical comment here: http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/papers/2003/n1480.html#309)

Also please note that those LWG members who in Copenhagen felt that "a sentry's constructor should not catch exceptions, because sentries should only be used within (un)formatted input functions and that exception handling is the responsibility of those functions, not of the sentries," as noted here http://anubis.dkuug.dk/jtc1/sc22/wg21/docs/papers/2001/n1310.html#309 would in effect be either arguing for the behavior described in [1] or for extractors implemented along the lines of [3].

The original proposed resolution (Revision 25 of the issues list) clarifies the role of the sentry ctor WRT exception handling by making it clear that extractors (both library or user-defined) should be implemented along the lines of [2] (as opposed to [1]) and that no exception thrown from the callees should propagate out of either function unless badbit is also set in exceptions().

[1] Extractor that catches exceptions thrown from sentry:

struct S { long i; };

istream& operator>> (istream &strm, S &s)
{
    ios::iostate err = ios::goodbit;
    try {
        const istream::sentry guard (strm, false);
        if (guard) {
            use_facet<num_get<char> >(strm.getloc ())
                .get (istreambuf_iterator<char>(strm),
                      istreambuf_iterator<char>(),
                      strm, err, s.i);
        }
    }
    catch (...) {
        bool rethrow;
        try {
            strm.setstate (ios::badbit);
            rethrow = false;
        }
        catch (...) {
            rethrow = true;
        }
        if (rethrow)
            throw;
    }
    if (err)
        strm.setstate (err);
    return strm;
}

[2] Extractor that propagates exceptions thrown from sentry:

istream& operator>> (istream &strm, S &s)
{
    istream::sentry guard (strm, false);
    if (guard) {
        ios::iostate err = ios::goodbit;
        try {
            use_facet<num_get<char> >(strm.getloc ())
                .get (istreambuf_iterator<char>(strm),
                      istreambuf_iterator<char>(),
                      strm, err, s.i);
        }
        catch (...) {
            bool rethrow;
            try {
                strm.setstate (ios::badbit);
                rethrow = false;
            }
            catch (...) {
                rethrow = true;
            }
            if (rethrow)
                throw;
        }
        if (err)
            strm.setstate (err);
    }
    return strm;
}

[3] Extractor that catches exceptions thrown from sentry but doesn't set badbit if the exception was thrown as a result of a call to strm.clear().

istream& operator>> (istream &strm, S &s)
{
    const ios::iostate state = strm.rdstate ();
    const ios::iostate except = strm.exceptions ();
    ios::iostate err = std::ios::goodbit;
    bool thrown = true;
    try {
        const istream::sentry guard (strm, false);
        thrown = false;
        if (guard) {
            use_facet<num_get<char> >(strm.getloc ())
                .get (istreambuf_iterator<char>(strm),
                      istreambuf_iterator<char>(),
                      strm, err, s.i);
        }
    }
    catch (...) {
        if (thrown && state & except)
            throw;
        try {
            strm.setstate (ios::badbit);
            thrown = false;
        }
        catch (...) {
            thrown = true;
        }
        if (thrown)
            throw;
    }
    if (err)
        strm.setstate (err);

    return strm;
}

Proposed resolution:

Rationale:

The LWG agrees there is minor variation between implementations, but believes that it doesn't matter. This is a rarely used corner case. There is no evidence that this has any commercial importance or that it causes actual portability problems for customers trying to write code that runs on multiple implementations.


313. set_terminate and set_unexpected question

Section: 18.6.3.3 [lib.terminate]  Status: NAD  Submitter: Judy Ward  Date: 3 Apr 2001

According to section 18.6.3.3 of the standard, std::terminate() is supposed to call the terminate_handler in effect immediately after evaluating the throw expression.

Question: what if the terminate_handler in effect is itself std::terminate?

For example:

  #include <exception>

  int main () {
      std::set_terminate(std::terminate);
      throw 5;
      return 0;
  }

Is the implementation allowed to go into an infinite loop?

I think the same issue applies to std::set_unexpected.

Proposed resolution:

Rationale:

Infinite recursion is to be expected: users who set the terminate handler to terminate are explicitly asking for terminate to call itself.


314. Is the stack unwound when terminate() is called?

Section: 18.6.3.3 [lib.terminate]  Status: NAD  Submitter: Detlef Vollmann  Date: 11 Apr 2001

The standard appears to contradict itself about whether the stack is unwound when the implementation calls terminate().

From 18.6.3.3p2:

Calls the terminate_handler function in effect immediately after evaluating the throw-expression (lib.terminate.handler), if called by the implementation [...]

So the stack is guaranteed not to be unwound.

But from 15.3p9:

[...]whether or not the stack is unwound before this call to terminate() is implementation-defined (except.terminate).

And 15.5.1 actually defines that in most cases the stack is unwound.

Proposed resolution:

Rationale:

There is definitely no contradiction between the core and library clauses; nothing in the core clauses says that stack unwinding happens after terminate is called. 18.6.3.3p2 does not say anything about when terminate() is called; it merely specifies which terminate_handler is used.


323. abs() overloads in different headers

Section: 26.5 [lib.c.math]  Status: NAD Future  Submitter: Dave Abrahams  Date: 4 June 2001

Currently the standard mandates the following overloads of abs():

    abs(long), abs(int) in <cstdlib>

    abs(float), abs(double), abs(long double) in <cmath>

    template<class T> T abs(const complex<T>&) in <complex>

    template<class T> valarray<T> abs(const valarray<T>&); in <valarray>

The problem is that having only some overloads visible of a function that works on "implicitly inter-convertible" types is dangerous in practice. The headers that get included at any point in a translation unit can change unpredictably during program development/maintenance. The wrong overload might be unintentionally selected.

Currently, there is nothing that mandates the simultaneous visibility of these overloads. Indeed, some vendors have begun fastidiously reducing dependencies among their (public) headers as a QOI issue: it helps people to write portable code by refusing to compile unless all the correct headers are #included.

The same issue may exist for other functions in the library.

Redmond: PJP reports that C99 adds two new kinds of abs: complex, and int_max_abs.

Related issue: 343.

Proposed resolution:

Rationale:

The programs that could potentially be broken by this situation are already fragile, and somewhat contrived: For example, a user-defined class that has conversion overloads both to long and to float. If x is a value of such a class, then abs(x) would give the long version if the user included <cstdlib>, the float version if the user included <cmath>, and would be diagnosed as ambiguous at compile time if the user included both headers. The LWG couldn't find an example of a program whose meaning would be changed (as opposed to changing it from well-formed to ill-formed) simply by adding another standard header.

Since the harm seems minimal, and there don't seem to be any simple and noninvasive solutions, this is being closed as NAD. It is marked as "Future" for two reasons. First, it might be useful to define an <all> header that would include all Standard Library headers. Second, we should at least make sure that future library extensions don't make this problem worse.


326. Missing typedef in moneypunct_byname

Section: 22.2.6.4 [lib.locale.moneypunct.byname]  Status: NAD  Submitter: Martin Sebor  Date: 05 Jul 2001

The definition of the moneypunct facet contains the typedefs char_type and string_type. Only one of these names, string_type, is defined in the derived facet, moneypunct_byname.

Proposed resolution:

For consistency with the numpunct facet, add a typedef for char_type to the definition of the moneypunct_byname facet in 22.2.6.4 [lib.locale.moneypunct.byname].

Rationale:

The absence of the typedef is irrelevant. Users can still access the typedef, because it is inherited from the base class.


330. Misleading "exposition only" value in class locale definition

Section: 22.1.1 [lib.locale]  Status: NAD  Submitter: Martin Sebor  Date: 15 Jul 2001

The "exposition only" value of the std::locale::none constant shown in the definition of class locale is misleading in that it on many systems conflicts with the value assigned to one if the LC_XXX constants (specifically, LC_COLLATE on AIX, LC_ALL on HP-UX, LC_CTYPE on Linux and SunOS). This causes incorrect behavior when such a constant is passed to one of the locale member functions that accept a locale::category argument and interpret it as either the C LC_XXX constant or a bitmap of locale::category values. At least three major implementations adopt the suggested value without a change and consequently suffer from this problem.

For instance, the following code will (presumably) incorrectly copy facets belonging to the collate category from the German locale on AIX:

  std::locale l (std::locale ("C"), "de_DE", std::locale::none);

Proposed resolution:

Rationale:

The LWG agrees that it may be difficult to implement locale member functions in such a way that they can take either category arguments or the LC_ constants defined in <cctype>. In light of this requirement (22.1.1.1.1 [lib.locale.category], paragraph 2), and in light of the requirement in the preceding paragraph that it is possible to combine category bitmask elements with bitwise operations, defining the category elements is delicate, particularly if an implementor is constrained to work with a preexisting C library. (Just using the existing LC_ constants would not work in general.) There's no set of "exposition only" values that could give library implementors proper guidance in such a delicate matter. The non-normative example we're giving is no worse than any other choice would be.

See issue 347.


332. Consider adding increment and decrement operators to std::fpos< T >

Section: 27.4.3 [lib.fpos]  Status: NAD  Submitter: PremAnand M. Rao  Date: 27 Aug 2001

Increment and decrement operators are missing from Table 88 -- Position type requirements in 27.4.3 [lib.fpos].

Proposed resolution:

Table 88 (section 27.4.3) -- Position type requirements be updated to include increment and decrement operators.

expression        return type     operational    note

++p               fpos&           p += O(1)
p++               fpos            { P tmp = p;
                                    ++p;
                                    return tmp; }
--p               fpos&           p -= O(1)
p--               fpos            { P tmp = p;
                                    --p;
                                    return tmp; }

Rationale:

The LWG believes this is a request for extension, not a defect report. Additionally, nobody saw a clear need for this extension; fpos is used only in very limited ways.


343. Unspecified library header dependencies

Section: 21 [lib.strings], 23 [lib.containers], 27 [lib.input.output]  Status: NAD Future  Submitter: Martin Sebor  Date: 09 Oct 2001

The synopses of the C++ library headers clearly show which names are required to be defined in each header. Since in order to implement the classes and templates defined in these headers declarations of other templates (but not necessarily their definitions) are typically necessary the standard in 17.4.4, p1 permits library implementers to include any headers needed to implement the definitions in each header.

For instance, although it is not explicitly specified in the synopsis of <string>, at the point of definition of the std::basic_string template the declaration of the std::allocator template must be in scope. All current implementations simply include <memory> from within <string>, either directly or indirectly, to bring the declaration of std::allocator into scope.

Additionally, however, some implementation also include <istream> and <ostream> at the top of <string> to bring the declarations of std::basic_istream and std::basic_ostream into scope (which are needed in order to implement the string inserter and extractor operators (21.3.7.9 [lib.string.io])). Other implementations only include <iosfwd>, since strictly speaking, only the declarations and not the full definitions are necessary.

Obviously, it is possible to implement <string> without actually providing the full definitions of all the templates std::basic_string uses (std::allocator, std::basic_istream, and std::basic_ostream). Furthermore, not only is it possible, doing so is likely to have a positive effect on compile-time efficiency.

But while it may seem perfectly reasonable to expect a program that uses the std::basic_string insertion and extraction operators to also explicitly include <istream> or <ostream>, respectively, it doesn't seem reasonable to also expect it to explicitly include <memory>. Since what's reasonable and what isn't is highly subjective one would expect the standard to specify what can and what cannot be assumed. Unfortunately, that isn't the case.

The examples below demonstrate the issue.

Example 1:

It is not clear whether the following program is complete:

#include <string>

extern std::basic_ostream<char> &strm;

int main () {
    strm << std::string ("Hello, World!\n");
}

or whether one must explicitly include <memory> or <ostream> (or both) in addition to <string> in order for the program to compile.

Example 2:

Similarly, it is unclear whether the following program is complete:

#include <istream>

extern std::basic_iostream<char> &strm;

int main () {
    strm << "Hello, World!\n";
}

or whether one needs to explicitly include <ostream>, and perhaps even other headers containing the definitions of other required templates:

#include <ios>
#include <istream>
#include <ostream>
#include <streambuf>

extern std::basic_iostream<char> &strm;

int main () {
    strm << "Hello, World!\n";
}

Example 3:

Likewise, it seems unclear whether the program below is complete:

#include <iterator>

bool foo (std::istream_iterator<int> a, std::istream_iterator<int> b)
{
    return a == b;
}

int main () { }

or whether one should be required to include <istream>.

There are many more examples that demonstrate this lack of a requirement. I believe that in a good number of cases it would be unreasonable to require that a program explicitly include all the headers necessary for a particular template to be specialized, but I think that there are cases such as some of those above where it would be desirable to allow implementations to include only as much as necessary and not more.

Proposed resolution:

For every C++ library header, supply a minimum set of other C++ library headers that are required to be included by that header. The proposed list is below (C++ headers for C Library Facilities, table 12 in 17.4.1.2, p3, are omitted):

+------------+--------------------+
| C++ header |required to include |
+============+====================+
|<algorithm> |                    |
+------------+--------------------+
|<bitset>    |                    |
+------------+--------------------+
|<complex>   |                    |
+------------+--------------------+
|<deque>     |<memory>            |
+------------+--------------------+
|<exception> |                    |
+------------+--------------------+
|<fstream>   |<ios>               |
+------------+--------------------+
|<functional>|                    |
+------------+--------------------+
|<iomanip>   |<ios>               |
+------------+--------------------+
|<ios>       |<streambuf>         |
+------------+--------------------+
|<iosfwd>    |                    |
+------------+--------------------+
|<iostream>  |<istream>, <ostream>|
+------------+--------------------+
|<istream>   |<ios>               |
+------------+--------------------+
|<iterator>  |                    |
+------------+--------------------+
|<limits>    |                    |
+------------+--------------------+
|<list>      |<memory>            |
+------------+--------------------+
|<locale>    |                    |
+------------+--------------------+
|<map>       |<memory>            |
+------------+--------------------+
|<memory>    |                    |
+------------+--------------------+
|<new>       |<exception>         |
+------------+--------------------+
|<numeric>   |                    |
+------------+--------------------+
|<ostream>   |<ios>               |
+------------+--------------------+
|<queue>     |<deque>             |
+------------+--------------------+
|<set>       |<memory>            |
+------------+--------------------+
|<sstream>   |<ios>, <string>     |
+------------+--------------------+
|<stack>     |<deque>             |
+------------+--------------------+
|<stdexcept> |                    |
+------------+--------------------+
|<streambuf> |<ios>               |
+------------+--------------------+
|<string>    |<memory>            |
+------------+--------------------+
|<strstream> |                    |
+------------+--------------------+
|<typeinfo>  |<exception>         |
+------------+--------------------+
|<utility>   |                    |
+------------+--------------------+
|<valarray>  |                    |
+------------+--------------------+
|<vector>    |<memory>            |
+------------+--------------------+

Rationale:

The portability problem is real. A program that works correctly on one implementation might fail on another, because of different header dependencies. This problem was understood before the standard was completed, and it was a conscious design choice.

One possible way to deal with this, as a library extension, would be an <all> header.


344. grouping + showbase

Section: 22.2.2 [lib.category.numeric]  Status: NAD  Submitter: Howard Hinnant  Date: 13 Oct 2001

When both grouping and showbase are active and the basefield is octal, does the leading 0 participate in the grouping or not? For example, should one format as: 0,123,456 or 0123,456?

An analogy can be drawn with hexadecimal. It appears that 0x123,456 is preferred over 0x,123,456. However, this analogy is not universally accepted to apply to the octal base. The standard is not clear on how to format (or parse) in this manner.

Proposed resolution:

Insert into 22.2.3.1.2 [lib.facet.numpunct.virtuals] paragraph 3, just before the last sentence:

The leading hexadecimal base specifier "0x" does not participate in grouping. The leading '0' octal base specifier may participate in grouping. It is unspecified if the leading '0' participates in formatting octal numbers. In parsing octal numbers, the implementation is encouraged to accept both the leading '0' participating in the grouping, and not participating (e.g. 0123,456 or 0,123,456).

Rationale:

The current behavior may be unspecified, but it's not clear that it matters. This is an obscure corner case, since grouping is usually intended for the benefit of humans and oct/hex prefixes are usually intended for the benefit of machines. There is not a strong enough consensus in the LWG for action.


348. Minor issue with std::pair operator<

Section: 20.2.2 [lib.pairs]  Status: NAD Future  Submitter: Andy Sawyer  Date: 23 Oct 2001

The current wording of 20.2.2 [lib.pairs] p6 precludes the use of operator< on any pair type which contains a pointer.

Proposed resolution:

In 20.2.2 [lib.pairs] paragraph 6, replace:

    Returns: x.first < y.first || (!(y.first < x.first) && x.second <
        y.second).

With:

    Returns: std::less<T1>()( x.first, y.first ) ||
             (!std::less<T1>()( y.first, x.first) && 
             std::less<T2>()( x.second, y.second ) )

Rationale:

This is an instance of a much more general problem. If we want operator< to translate to std::less for pairs of pointers, where do we draw the line? The same issue applies to individual pointers, smart pointer wrappers, std::vector<T*>, and so on.

Andy Koenig suggests that the real issue here is that we aren't distinguishing adequately between two different orderings, a "useful ordering" and a "canonical ordering" that's used just because we sometimes need some ordering without caring much which ordering it is. Another example of the later is typeinfo's before.


350. allocator<>::address

Section: 20.4.1.1 [lib.allocator.members], 20.1.5 [lib.allocator.requirements], 17.4.1.1 [lib.contents]  Status: NAD Future  Submitter: Nathan Myers  Date: 25 Oct 2001

See c++std-lib-9006 and c++std-lib-9007. This issue is taken verbatim from -9007.

The core language feature allowing definition of operator&() applied to any non-builtin type makes that operator often unsafe to use in implementing libraries, including the Standard Library. The result is that many library facilities fail for legal user code, such as the fragment

  class A { private: A* operator&(); };
  std::vector<A> aa;

  class B { };
  B* operator&(B&) { return 0; }
  std::vector<B> ba;

In particular, the requirements table for Allocator (Table 32) specifies no semantics at all for member address(), and allocator<>::address is defined in terms of unadorned operator &.

Proposed resolution:

In 20.4.1.1, Change the definition of allocator<>::address from:

Returns: &x

to:

Returns: The value that the built in operator&(x) would return if not overloaded.

In 20.1.5, Table 32, add to the Notes column of the a.address(r) and a.address(s) lines, respectively:

  allocator<T>::address(r)
  allocator<T>::address(s)

In addition, in clause 17.4.1.1, add a statement:

The Standard Library does not apply operator& to any type for which operator& may be overloaded.

Rationale:

The LWG believes both examples are ill-formed. The contained type is required to be CopyConstructible (20.1.3 [lib.copyconstructible]), and that includes the requirement that &t return the usual types and values. Since allocators are intended to be used in conjunction with containers, and since the CopyConstructible requirements appear to have been written to deal with the concerns of this issue, the LWG feels it is NAD unless someone can come up with a well-formed example exhibiting a problem.

It may well be that the CopyConstructible requirements are too restrictive and that either the container requirements or the CopyConstructive requirements should be relaxed, but that's a far larger issue. Marking this issue as "future" as a pointer to that larger issue.


351. unary_negate and binary_negate: struct or class?

Section: 20.3 [lib.function.objects]  Status: NAD  Submitter: Dale Riley  Date: 12 Nov 2001

In 20.3 [lib.function.objects] the header <functional> synopsis declares the unary_negate and binary_negate function objects as struct. However in 20.3.5 [lib.negators] the unary_negate and binary_negate function objects are defined as class. Given the context, they are not "basic function objects" like negate, so this is either a typo or an editorial oversight.

[Taken from comp.std.c++]

Proposed resolution:

Change the synopsis to reflect the useage in 20.3.5 [lib.negators]

[Curaçao: Since the language permits "struct", the LWG views this as NAD. They suggest, however, that the Project Editor might wish to make the change as editorial.]


353. std::pair missing template assignment

Section: 20.2.2 [lib.pairs]  Status: NAD Future  Submitter: Martin Sebor  Date: 2 Dec 2001

The class template std::pair defines a template ctor (20.2.2, p4) but no template assignment operator. This may lead to inefficient code since assigning an object of pair<C, D> to pair<A, B> where the types C and D are distinct from but convertible to A and B, respectively, results in a call to the template copy ctor to construct an unnamed temporary of type pair<A, B> followed by an ordinary (perhaps implicitly defined) assignment operator, instead of just a straight assignment.

Proposed resolution:

Add the following declaration to the definition of std::pair:

    template<class U, class V>
    pair& operator=(const pair<U, V> &p);

And also add a paragraph describing the effects of the function template to the end of 20.2.2:

    template<class U, class V>
    pair& operator=(const pair<U, V> &p);

Effects: first = p.first; second = p.second; Returns: *this

[Curaçao: There is no indication this is was anything other than a design decision, and thus NAD.  May be appropriate for a future standard.]


356. Meaning of ctype_base::mask enumerators

Section: 22.2.1 [lib.category.ctype]  Status: NAD  Submitter: Matt Austern  Date: 23 Jan 2002

What should the following program print?

  #include <locale>
  #include <iostream>

  class my_ctype : public std::ctype<char>
  {
    typedef std::ctype<char> base;
  public:
    my_ctype(std::size_t refs = 0) : base(my_table, false, refs)
    {
      std::copy(base::classic_table(), base::classic_table() + base::table_size,
                my_table);
      my_table[(unsigned char) '_'] = (base::mask) (base::print | base::space);
    }
  private:
    mask my_table[base::table_size];
  };

  int main()
  {
    my_ctype ct;
    std::cout << "isspace: " << ct.is(std::ctype_base::space, '_') << "    "
              << "isalpha: " << ct.is(std::ctype_base::alpha, '_') << std::endl;
  }

The goal is to create a facet where '_' is treated as whitespace.

On gcc 3.0, this program prints "isspace: 1 isalpha: 0". On Microsoft C++ it prints "isspace: 1 isalpha: 1".

I believe that both implementations are legal, and the standard does not give enough guidance for users to be able to use std::ctype's protected interface portably.

The above program assumes that ctype_base::mask enumerators like space and print are disjoint, and that the way to say that a character is both a space and a printing character is to or those two enumerators together. This is suggested by the "exposition only" values in 22.2.1 [lib.category.ctype], but it is nowhere specified in normative text. An alternative interpretation is that the more specific categories subsume the less specific. The above program gives the results it does on the Microsoft compiler because, on that compiler, print has all the bits set for each specific printing character class.

From the point of view of std::ctype's public interface, there's no important difference between these two techniques. From the point of view of the protected interface, there is. If I'm defining a facet that inherits from std::ctype<char>, I'm the one who defines the value that table()['a'] returns. I need to know what combination of mask values I should use. This isn't so very esoteric: it's exactly why std::ctype has a protected interface. If we care about users being able to write their own ctype facets, we have to give them a portable way to do it.

Related reflector messages: lib-9224, lib-9226, lib-9229, lib-9270, lib-9272, lib-9273, lib-9274, lib-9277, lib-9279.

Issue 339 is related, but not identical. The proposed resolution if issue 339 says that ctype_base::mask must be a bitmask type. It does not say that the ctype_base::mask elements are bitmask elements, so it doesn't directly affect this issue.

More comments from Benjamin Kosnik, who believes that that C99 compatibility essentially requires what we're calling option 1 below.

I think the C99 standard is clear, that isspace -> !isalpha.
--------

#include <locale>
#include <iostream>

class my_ctype : public std::ctype<char>
{
private:
  typedef std::ctype<char> base;
  mask my_table[base::table_size];

public:
  my_ctype(std::size_t refs = 0) : base(my_table, false, refs)
  {
    std::copy(base::classic_table(), base::classic_table() + base::table_size,
              my_table);
    mask both = base::print | base::space;
    my_table[static_cast<mask>('_')] = both;
  }
};

int main()
{
  using namespace std;
  my_ctype ct;
  cout << "isspace: " << ct.is(ctype_base::space, '_') << endl;
  cout << "isprint: " << ct.is(ctype_base::print, '_') << endl;

  // ISO C99, isalpha iff upper | lower set, and !space.
  // 7.5, p 193
  // -> looks like g++ behavior is correct.
  // 356 -> bitmask elements are required for ctype_base
  // 339 -> bitmask type required for mask
  cout << "isalpha: " << ct.is(ctype_base::alpha, '_') << endl;
}

Proposed resolution:

Informally, we have three choices:

  1. Require that the enumerators are disjoint (except for alnum and graph)
  2. Require that the enumerators are not disjoint, and specify which of them subsume which others. (e.g. mandate that lower includes alpha and print)
  3. Explicitly leave this unspecified, which the result that the above program is not portable.

Either of the first two options is just as good from the standpoint of portability. Either one will require some implementations to change.

Rationale:

The LWG agrees that this is a real ambiguity, and that both interpretations are conforming under the existing standard. However, there's no evidence that it's causing problems for real users. Users who want to define ctype facets portably can test the ctype_base masks to see which interpretation is being used.


357. <cmath> float functions cannot return HUGE_VAL

Section: 26.5 [lib.c.math]  Status: NAD  Submitter: Ray Lischner  Date: 26 Feb 2002

The float versions of the math functions have no meaningful value to return for a range error. The long double versions have a value they can return, but it isn't necessarily the most reasonable value.

Section 26.5 [lib.c.math], paragraph 5, says that C++ "adds float and long double overloaded versions of these functions, with the same semantics," referring to the math functions from the C90 standard.

The C90 standard, in section 7.5.1, paragraph 3, says that functions return "the value of the macro HUGE_VAL" when they encounter a range error. Section 7.5, paragraph 2, defines HUGE_VAL as a macro that "expands to a positive double expression, not necessarily representable as a float."

Therefore, the float versions of the math functions have no way to signal a range error. [Curaçao: The LWG notes that this isn't strictly correct, since errno is set.] The semantics require that they return HUGE_VAL, but they cannot because HUGE_VAL might not be representable as a float.

The problem with long double functions is less severe because HUGE_VAL is representable as a long double. On the other hand, it might not be a "huge" long double value, and might fall well within the range of normal return values for a long double function. Therefore, it does not make sense for a long double function to return a double (HUGE_VAL) for a range error.

Proposed resolution:

Curaçao: C99 was faced with a similar problem, which they fixed by adding HUGE_VALF and HUGE_VALL in addition to HUGE_VAL.

C++ must also fix, but it should be done in the context of the general C99 based changes to C++, not via DR. Thus the LWG in Curaçao felt the resolution should be NAD, FUTURE, but the issue is being held open for one more meeting to ensure LWG members not present during the discussion concur.

Rationale:

Will be fixed as part of more general work in the TR.


361. num_get<>::do_get (..., void*&) checks grouping

Section: 22.2.2.2.2 [lib.facet.num.put.virtuals]  Status: NAD  Submitter: Martin Sebor  Date: 12 Mar 2002

22.2.2.2.2, p12 specifies that thousands_sep is to be inserted only for integral types (issue 282 suggests that this should be done for all arithmetic types).

22.2.2.1.2, p12 requires that grouping be checked for all extractors including that for void*.

I don't think that's right. void* values should not be checked for grouping, should they? (Although if they should, then num_put needs to write them out, otherwise their extraction will fail.)

Proposed resolution:

Change the first sentence of 22.2.2.2.2, p12 from

Digit grouping is checked. That is, the positions of discarded separators is examined for consistency with use_facet<numpunct<charT> >(loc).grouping(). If they are not consistent then ios_base::failbit is assigned to err.

to

Except for conversions to void*, digit grouping is checked...

Rationale:

This would be a change: as it stands, the standard clearly specifies that grouping applies to void*. A survey of existing practice shows that most existing implementations do that, as they should.


366. Excessive const-qualification

Section: 27 [lib.input.output]  Status: NAD  Submitter: Walter Brown, Marc Paterno  Date: 10 May 2002

The following member functions are declared const, yet return non-const pointers. We believe they are should be changed, because they allow code that may surprise the user. See document N1360 for details and rationale.

[Santa Cruz: the real issue is that we've got const member functions that return pointers to non-const, and N1360 proposes replacing them by overloaded pairs. There isn't a consensus about whether this is a real issue, since we've never said what our constness policy is for iostreams. N1360 relies on a distinction between physical constness and logical constness; that distinction, or those terms, does not appear in the standard.]

Proposed resolution:

In 27.4.4 and 27.4.4.2

Replace

  basic_ostream<charT,traits>* tie() const;

with

  basic_ostream<charT,traits>* tie();
  const basic_ostream<charT,traits>* tie() const;

and replace

  basic_streambuf<charT,traits>* rdbuf() const;

with

  basic_streambuf<charT,traits>* rdbuf();
  const basic_streambuf<charT,traits>* rdbuf() const;

In 27.5.2 and 27.5.2.3.1

Replace

  char_type* eback() const;

with

  char_type* eback();
  const char_type* eback() const;

Replace

  char_type gptr() const;

with

  char_type* gptr();
  const char_type* gptr() const;

Replace

  char_type* egptr() const;

with

  char_type* egptr();
  const char_type* egptr() const;

In 27.5.2 and 27.5.2.3.2

Replace

  char_type* pbase() const;

with

  char_type* pbase();
  const char_type* pbase() const;

Replace

  char_type* pptr() const;

with

  char_type* pptr();
  const char_type* pptr() const;

Replace

  char_type* epptr() const;

with

  char_type* epptr();
  const char_type* epptr() const;

In 27.7.2, 27.7.2.2, 27.7.3 27.7.3.2, 27.7.4, and 27.7.6

Replace

  basic_stringbuf<charT,traits,Allocator>* rdbuf() const;

with

  basic_stringbuf<charT,traits,Allocator>* rdbuf();
  const basic_stringbuf<charT,traits,Allocator>* rdbuf() const;

In 27.8.1.5, 27.8.1.7, 27.8.1.8, 27.8.1.10, 27.8.1.11, and 27.8.1.13

Replace

  basic_filebuf<charT,traits>* rdbuf() const;

with

  basic_filebuf<charT,traits>* rdbuf();
  const basic_filebuf<charT,traits>* rdbuf() const;

Rationale:

The existing specification is a bit sloppy, but there's no particular reason to change this other than tidiness, and there are a number of ways in which streams might have been designed differently if we were starting today. There's no evidence that the existing constness policy is harming users. We might consider a different constness policy as part of a full stream redesign.


367. remove_copy/remove_copy_if and Input Iterators

Section: 25.2.7 [lib.alg.remove]  Status: NAD  Submitter: Anthony Williams  Date: 13 May 2002

remove_copy and remove_copy_if (25.2.7 [lib.alg.remove]) permit their input range to be marked with Input Iterators. However, since two operations are required against the elements to copy (comparison and assigment), when the input range uses Input Iterators, a temporary copy must be taken to avoid dereferencing the iterator twice. This therefore requires the value type of the InputIterator to be CopyConstructible. If the iterators are at least Forward Iterators, then the iterator can be dereferenced twice, or a reference to the result maintained, so the temporary is not required.

Proposed resolution:

Add "If InputIterator does not meet the requirements of forward iterator, then the value type of InputIterator must be copy constructible. Otherwise copy constructible is not required." to 25.2.7 [lib.alg.remove] paragraph 6.

Rationale:

The assumption is that an input iterator can't be dereferenced twice. There's no basis for that assumption in the Standard.


368. basic_string::replace has two "Throws" paragraphs

Section: 21.3.5.6 [lib.string::replace]  Status: NAD  Submitter: Beman Dawes  Date: 3 Jun 2002

21.3.5.6 [lib.string::replace] basic_string::replace, second signature, given in paragraph 1, has two "Throws" paragraphs (3 and 5).

In addition, the second "Throws" paragraph (5) includes specification (beginning with "Otherwise, the function replaces ...") that should be part of the "Effects" paragraph.

Proposed resolution:

Rationale:

This is editorial. Both "throws" statements are true. The bug is just that the second one should be a sentence, part of the "Effects" clause, not a separate "Throws". The project editor has been notified.


372. Inconsistent description of stdlib exceptions

Section: 17.4.4.8 [lib.res.on.exception.handling], 18.6.1 [lib.exception],   Status: NAD  Submitter: Randy Maddox  Date: 22 Jul 2002

Paragraph 3 under clause 17.4.4.8 [lib.res.on.exception.handling], Restrictions on Exception Handling, states that "Any other functions defined in the C++ Standard Library that do not have an exception-specification may throw implementation-defined exceptions unless otherwise specified." This statement is followed by a reference to footnote 178 at the bottom of that page which states, apparently in reference to the C++ Standard Library, that "Library implementations are encouraged (but not required) to report errors by throwing exceptions from (or derived from) the standard exceptions."

These statements appear to be in direct contradiction to clause 18.6.1 [lib.exception], which states "The class exception defines the base class for the types of objects thrown as exceptions by the C++ Standard library components ...".

Is this inconsistent?

Proposed resolution:

Rationale:

Clause 17 is setting the overall library requirements, and it's clear and consistent. This sentence from Clause 18 is descriptive, not setting a requirement on any other class.


374. moneypunct::frac_digits returns int not unsigned

Section: 22.2.6.3.1 [lib.locale.moneypunct.members], 22.2.6.3.2 [lib.locale.moneypunct.virtuals]  Status: NAD  Submitter: Ray Lischner  Date: 8 Aug 2002

In section 22.2.6.3.1 [lib.locale.moneypunct.members], frac_digits() returns type "int". This implies that frac_digits() might return a negative value, but a negative value is nonsensical. It should return "unsigned".

Similarly, in section 22.2.6.3.2 [lib.locale.moneypunct.virtuals], do_frac_digits() should return "unsigned".

Proposed resolution:

Rationale:

Regardless of whether the return value is int or unsigned, it's always conceivable that frac_digits might return a nonsensical value. (Is 4294967295 really any better than -1?) The clients of moneypunct, the get and put facets, can and do perform range checks.


377. basic_string::insert and length_error

Section: 21.3.5.4 [lib.string::insert]  Status: NAD  Submitter: Ray Lischner  Date: 16 Aug 2002

Section 21.3.5.4 [lib.string::insert], paragraph 4, contains the following, "Then throws length_error if size() >= npos - rlen."

Related to DR 83, this sentence should probably be removed.

Proposed resolution:

Rationale:

This requirement is redundant but correct. No change is needed.


378. locale immutability and locale::operator=()

Section: 22.1.1 [lib.locale]  Status: Dup  Submitter: Martin Sebor  Date: 6 Sep 2002

I think there is a problem with 22.1.1, p6 which says that

    -6- An instance of locale is immutable; once a facet reference
        is obtained from it, that reference remains usable as long
        as the locale value itself exists.

and 22.1.1.2, p4:

    const locale& operator=(const locale& other) throw();

    -4- Effects: Creates a copy of other, replacing the current value.

How can a reference to a facet obtained from a locale object remain valid after an assignment that clearly must replace all the facets in the locale object? Imagine a program such as this

    std::locale loc ("de_DE");
    const std::ctype<char> &r0 = std::use_facet<std::ctype<char> >(loc);
    loc = std::locale ("en_US");
    const std::ctype<char> &r1 = std::use_facet<std::ctype<char> >(loc);

Is r0 really supposed to be preserved and destroyed only when loc goes out of scope?

Proposed resolution:

[Summer '04 mid-meeting mailing: Martin and Dietmar believe this is a duplicate of issue 31 and recommend that it be closed. ]


388. Use of complex as a key in associative containers

Section: 26.2 [lib.complex.numbers]  Status: NAD Future  Submitter: Gabriel Dos Reis  Date: 8 Nov 2002

Practice with std::complex<> and the associative containers occasionally reveals artificial and distracting issues with constructs resembling: std::set<std::complex<double> > s;

The main reason for the above to fail is the absence of an approriate definition for std::less<std::complex<T> >. That in turn comes from the definition of the primary template std::less<> in terms of operator<.

The usual argument goes as follows: Since there is no ordering over the complex field compatible with field operations it makes little sense to define a function operator< operating on the datatype std::complex<T>. That is fine. However, that reasoning does not carry over to std::less<T> which is used, among other things, by associative containers as an ordering useful to meet complexity requirements.

Related issue: 348.

Proposed resolution:

Informally: Add a specialization of std::less for std::complex.

Rationale:

Discussed in Santa Cruz. An overwhelming majority of the LWG believes this should not be treated a DR: it's a request for a design change, not a defect in the existing standard. Most people (10-3) believed that we probably don't want this change, period: as with issue 348, it's hard to know where to draw the line. The LWG noted that users who want to put objects into an associative container for which operator< isn't defined can simply provide their own comparison function object.


390. CopyConstructible requirements too strict

Section: 20.1.3 [lib.copyconstructible]  Status: NAD Future  Submitter: Doug Gregor  Date: 24 Oct 2002

The CopyConstructible requirements in Table 30 state that for an object t of type T (where T is CopyConstructible), the expression &t returns the address of t (with type T*). This requirement is overly strict, in that it disallows types that overload operator& to not return a value of type T*. This occurs, for instance, in the Boost.Lambda library, where operator& is overloaded for a Boost.Lambda function object to return another function object.

Example:

  std::vector<int> u, v;
  int x;
  // ...
  std::transform(u.begin(), u.end(), std::back_inserter(v), _1 * x);

_1 * x returns an unnamed function object with operator& overloaded to not return T* , therefore rendering the std::transform call ill-formed. However, most standard library implementations will compile this code properly, and the viability of such binder libraries is severely hindered by the unnecessary restriction in the CopyConstructible requirements.

For reference, the address of an object can be retrieved without using the address-of operator with the following function template:

  template <typename T> T* addressof(T& v)
  {
    return reinterpret_cast<T*>(
         &const_cast<char&>(reinterpret_cast<const volatile char &>(v)));
  }

Note: this relates directly to library issue 350, which will need to be reexamined if the CopyConstructible requirements change.

Proposed resolution:

Remove the last two rows of Table 30, eliminating the requirements that &t and &u return the address of t and u, respectively.

Rationale:

This was a deliberate design decision. Perhaps it should be reconsidered for C++0x.


392. 'equivalence' for input iterators

Section: 24.1.1 [lib.input.iterators]  Status: NAD  Submitter: Corwin Joy  Date: 11 Dec 2002

In section 24.1.1 [lib.input.iterators] table 72 - 'Input Iterator Requirements' we have as a postcondition of *a: "If a==b and (a, b) is in the domain of == then *a is equivalent to *b".

In section 24.5.3.5 [lib.istreambuf.iterator::equal] it states that "istreambuf_iterator::equal returns true if and only if both iterators are at end-of-stream, or neither is at end-of-stream, regardless of what streambuf object they use." (My emphasis).

The defect is that either 'equivalent' needs to be more precisely defined or the conditions for equality in 24.5.3.5 [lib.istreambuf.iterator::equal] are incorrect. (Or both).

Consider the following example:

   #include <iostream>
   #include <fstream>
   #include <iterator>
   using namespace std;

   int main() {
    ifstream file1("file1.txt"), file2("file2.txt");
    istreambuf_iterator<char> f1(file1), f2(file2);
    cout << "f1 == f2 : " << boolalpha << (f1 == f2) << endl;
    cout << "f1 = " << *f1 << endl;
    cout << "f2 = " << *f2 << endl;
    return 0;
   }

Now assuming that neither f1 or f2 are at the end-of-stream then f1 == f2 by 24.5.3.5 [lib.istreambuf.iterator::equal].

However, it is unlikely that *f1 will give the same value as *f2 except by accident.

So what does *f1 'equivalent' to *f2 mean? I think the standard should be clearer on this point, or at least be explicit that this does not mean that *f1 and *f2 are required to have the same value in the case of input iterators.

Proposed resolution:

Rationale:

The two iterators aer not in the domain of ==


399. volations of unformatted input function requirements

Section: 27.6.1.3 [lib.istream.unformatted]  Status: NAD  Submitter: Martin Sebor  Date: 5 Jan 2003

The Effects clauses for the two functions below violate the general requirements on unformatted input functions outlined in 27.6.1.3: they do not begin by constructing a sentry object. Instead, they begin by calling widen ('\n'), which may throw an exception. The exception is then allowed to propagate from the unformatted input function irrespective of the setting of exceptions().

Note that in light of 27.6.1.1, p3 and p4, the fact that the functions allow exceptions thrown from widen() to propagate may not strictly speaking be a defect (but the fact that the functions do not start by constructing a sentry object still is). However, since an exception thrown from ctype<charT> ::widen() during any other input operation (say, from within a call to num_get<charT>::get()) will be caught and cause badbit to be set, these two functions should not be treated differently for the sake of consistency.

Proposed resolution:

Rationale:

Not a defect. The standard is consistent, and the behavior required by the standard is unambiguous. Yes, it's theoretically possible for widen to throw. (Not that this will happen for the default ctype facet or for most real-world replacement ctype facets.) Users who define ctype facets that can throw, and who care about this behavior, can use alternative signatures that don't call widen.


429. typo in basic_ios::clear(iostate)

Section: 27.4.4.3 [lib.iostate.flags]  Status: Dup  Submitter: Martin Sebor  Date: 18 Sep 2003

The Effects clause in 27.4.4.3, p5 describing the effects of a call to the ios_base member function clear(iostate state) says that the function only throws if the respective bits are already set prior to the function call. That's obviously not the intent. If it was, a call to clear(badbit) on an object for which (rdstate() == goodbit && exceptions() == badbit) holds would not result in an exception being thrown.

Proposed resolution:

The text ought to be changed from
"If (rdstate() & exceptions()) == 0, returns. ..."
to
"If (state & exceptions()) == 0, returns. ..."

Rationale:

This is a duplicate of issue 412.


433. Contradiction in specification of unexpected()

Section: 18.6.2.4 [lib.unexpected]  Status: NAD  Submitter: Vyatcheslav Sysoltsev  Date: 29 Sep 2003

Clause 15.5.2 [except.unexpected] paragraph 1 says that "void unexpected(); is called (18.6.2) immediately after completing the stack unwinding for the former function", but 18.6.2.4 (Effects) says that "void unexpected(); . . . Calls the unexpected_handler function in effect immediately after evaluating the throwexpression (18.6.2.2),". Isn't here a contradiction: 15.5.2 requires stack have been unwound when in void unexpected() and therefore in unexpected_handler but 18.6.2.4 claims that unexpected_handler is called "in effect immediately" after evaluation of throw expression is finished, so there is no space left for stack to be unwound therefore? I think the phrase "in effect immediately" should be removed from the standard because it brings ambiguity in understanding.

Proposed resolution:

Rationale:

There is no contradiction. The phrase "in effect immediately" is just to clarify which handler is to be called.


437. Formatted output of function pointers is confusing

Section: 27.6.2.5.2 [lib.ostream.inserters.arithmetic]  Status: NAD  Submitter: Ivan Godard  Date: 24 Oct 2003

Given:

void f(int) {}
void(*g)(int) = f;
cout << g;

(with the expected #include and usings), the value printed is a rather surprising "true". Rather useless too.

The standard defines:

ostream& operator<<(ostream&, void*);

which picks up all data pointers and prints their hex value, but does not pick up function pointers because there is no default conversion from function pointer to void*. Absent that, we fall back to legacy conversions from C and the function pointer is converted to bool.

There should be an analogous inserter that prints the address of a function pointer.

Proposed resolution:

Rationale:

This is indeed a wart, but there is no good way to solve it. C doesn't provide a portable way of outputting the address of a function point either.


439. Should facets be copyable?

Section: 22.2 [lib.locale.categories]  Status: NAD  Submitter: Matt Austern  Date: 2 Nov 2003

The following facets classes have no copy constructors described in the standard, which, according to the standard, means that they are supposed to use the compiler-generated defaults. Default copy behavior is probably inappropriate. We should either make these classes uncopyable or else specify exactly what their constructors do.

Related issue: 421.

        ctype_base
        ctype
        ctype_byname
        ctype<char>
        ctype_byname<char>
        codecvt_base
        codecvt
        codecvt_byname
        num_get
        num_put
        numpunct
        numpunct_byname
        collate
        collate_byname
        time_base
        time_get
        time_get_byname
        time_put
        time_put_byname
        money_get
        money_put
        money_base
        moneypunct
        moneypunct_byname
        messages_base
        messages
        messages_byname

Proposed resolution:

Rationale:

The copy constructor in the base class is private.


440. Should std::complex use unqualified transcendentals?

Section: 26.2.8 [lib.complex.transcendentals]  Status: NAD  Submitter: Matt Austern  Date: 5 Nov 2003

Operations like pow and exp on complex<T> are typically implemented in terms of operations like sin and cos on T. Should implementations write this as std::sin, or as plain unqualified sin?

The issue, of course, is whether we want to use argument-dependent lookup in the case where T is a user-defined type. This is similar to the issue of valarray transcendentals, as discussed in issue 226.

This issue differs from valarray transcendentals in two important ways. First, "the effect of instantiating the template complex for types other than float, double or long double is unspecified." (26.2.1 [lib.complex.synopsis]) Second, the standard does not dictate implementation, so there is no guarantee that a particular real math function is used in the implementation of a particular complex function.

Proposed resolution:

Rationale:

If you instantiate std::complex for user-defined types, all bets are off.


447. Wrong template argument for time facets

Section: 22.1.1.1.1 [lib.locale.category]  Status: Dup  Submitter: Pete Becker  Date: 26 Dec 2003

22.1.1.1.1/4, table 52, "Required Instantiations", lists, among others:

    time_get<char,InputIterator>
    time_get_byname<char,InputIterator>
    time_get<wchar_t,OutputIterator>
    time_get_byname<wchar_t,OutputIterator>

The second argument to the last two should be InputIterator, not OutputIterator.

Proposed resolution:

Change the second template argument to InputIterator.

Rationale:

Duplicate of issue 327

450. set::find is inconsistent with associative container requirements

Section: 23.3.3 [lib.set]  Status: Dup  Submitter: Bill Plauger  Date: 30 Jan 2004

map/multimap have:

	iterator find(const key_type& x) const;
	const_iterator find(const key_type& x) const;

which is consistent with the table of associative container requirements. But set/multiset have:

	iterator find(const key_type&) const;

set/multiset should look like map/multimap, and honor the requirements table, in this regard.

Proposed resolution:

Rationale:

Duplicate of issue 214.


451. Associative erase should return an iterator

Section: 23.1.2 [lib.associative.reqmts], 23.3 [lib.associative]  Status: Dup  Submitter: Bill Plauger  Date: 30 Jan 2004

map/multimap/set/multiset have:

	void erase(iterator);
	void erase(iterator, iterator);

But there's no good reason why these can't return an iterator, as for vector/deque/list:

	iterator erase(iterator);
	iterator erase(iterator, iterator);

Proposed resolution:

Informally: The table of associative container requirements, and the relevant template classes, should return an iterator designating the first element beyond the erased subrange.

Rationale:

Duplicate of 130


452.  locale::combine should be permitted to generate a named locale

Section: 22.1.1.3 [lib.locale.members]  Status: NAD  Submitter: Bill Plauger  Date: 30 Jan 2004

template<class Facet>
	locale::combine(const locale&) const;

is obliged to create a locale that has no name. This is overspecification and overkill. The resulting locale should follow the usual rules -- it has a name if the locale argument has a name and Facet is one of the standard facets.

[ Sydney and post-Sydney (see c++std-lib-13439, c++std-lib-13440, c++std-lib-13443): agreed that it's overkill to say that the locale is obligated to be nameless. However, we also can't require it to have a name. At the moment, locale names are based on categories and not on individual facets. If a locale contains two different facets of different names from the same category, then this would not fit into existing naming schemes. We need to give implementations more freedom. Bill will provide wording. ]

Proposed resolution:

Rationale:

After further discussion the LWG decided to close this as NAD. The fundamental problem is that names right now are per-category, not per-facet. The combine member function works at the wrong level of granularity.


472. Missing "Returns" clause in std::equal_range

Section: 25.3.3.3 [lib.equal.range]  Status: Dup  Submitter: Prateek R Karandikar  Date: 29 Feb 1900

There is no "Returns:" clause for std::equal_range, which returns non-void.

Proposed resolution:

Rationale:

Fixed as part of issue 270.


476. Forward Iterator implied mutability

Section: 24.1.3 [lib.forward.iterators]  Status: NAD  Submitter: Dave Abrahams  Date: 9 Jul 2004

24.1/3 says:

Forward iterators satisfy all the requirements of the input and output iterators and can be used whenever either kind is specified

The problem is that satisfying the requirements of output iterator means that you can always assign *something* into the result of dereferencing it. That makes almost all non-mutable forward iterators non-conforming. I think we need to sever the refinement relationship between forward iterator and output iterator.

Related issue: 200. But this is not a dup.

Proposed resolution:

Rationale:

Yes, 24.1/3 does say that. But it's introductory material. The precise specification is in 24.1.3, and the requrements table there is right. We don't need to fine-tune introductory wording. (Especially since this wording is likely to be changed as part of the iterator overhaul.)


477. Operator-> for const forward iterators

Section: 24.1.3 [lib.forward.iterators]  Status: Dup  Submitter: Dave Abrahams  Date: 11 Jul 2004

The Forward Iterator requirements table contains the following:

 expression  return type         operational  precondition
                                  semantics
  ==========  ==================  ===========  ==========================
  a->m        U& if X is mutable, (*a).m       pre: (*a).m is well-defined.
              otherwise const U&

  r->m        U&                  (*r).m       pre: (*r).m is well-defined.

The first line is exactly right. The second line is wrong. Basically it implies that the const-ness of the iterator affects the const-ness of referenced members. But Paragraph 11 of [lib.iterator.requirements] says:

In the following sections, a and b denote values of type const X, n denotes a value of the difference type Distance, u, tmp, and m denote identifiers, r denotes a value of X&, t denotes a value of value type T, o denotes a value of some type that is writable to the output iterator.

AFAICT if we need the second line at all, it should read the same as the first line.

Related issue: 478

Proposed resolution:

Rationale:

The LWG agrees that this is a real problem. Marked as a DUP because the LWG chose to adopt the solution proposed in 478.


480. unary_function and binary_function should have protected nonvirtual destructors

Section: 20.3.1 [lib.base]  Status: NAD  Submitter: Joe Gottman  Date: 19 Aug 2004

The classes std::unary_function and std::binary_function are both designed to be inherited from but contain no virtual functions. This makes it too easy for a novice programmer to write code like binary_function<int, int, int> *p = new plus<int>; delete p;

There are two common ways to prevent this source of undefined behavior: give the base class a public virtual destructor, or give it a protected nonvirtual destructor. Since unary_function and binary_function have no other virtual functions, (note in particular the absence of an operator()() ), it would cost too much to give them public virtual destructors. Therefore, they should be given protected nonvirtual destructors.

Proposed resolution:

Change Paragraph 20.3.1 of the Standard from

    template <class Arg, class Result>
    struct unary_function {
        typedef Arg argument_type;
        typedef Result result_type;
    };

    template <class Arg1, class Arg2, class Result>
    struct binary_function {
        typedef Arg1 first_argument_type;
        typedef Arg2 second_argument_type;
        typedef Result result_type;
    };

to

    template <class Arg, class Result>
        struct unary_function {
        typedef Arg argument_type;
        typedef Result result_type;
    protected:
        ~unary_function() {}
    };

    template <class Arg1, class Arg2, class Result>
    struct binary_function {
        typedef Arg1 first_argument_type;
        typedef Arg2 second_argument_type;
        typedef Result result_type;
    protected:
        ~binary_function() {}
    };

Rationale:

The LWG doesn't believe the existing definition causes anybody any concrete harm.


481. unique's effects on the range [result, last)

Section: 25.2.8 [lib.alg.unique]  Status: NAD  Submitter: Andrew Koenig  Date: 30 Aug 2004

The standard says that unique(first, last) "eliminates all but the first element from every consecutive group of equal elements" in [first, last) and returns "the end of the resulting range". So a postcondition is that [first, result) is the same as the old [first, last) except that duplicates have been eliminated.

What postconditions are there on the range [result, last)? One might argue that the standard says nothing about those values, so they can be anything. One might also argue that the standard doesn't permit those values to be changed, so they must not be. Should the standard say something explicit one way or the other?

Proposed resolution:

Rationale:

We don't want to make many guarantees about what's in [result, end). Maybe we aren't being quite explicit enough about not being explicit, but it's hard to think that's a major problem.


483. Heterogeneous equality and EqualityComparable

Section: 25.1 [lib.alg.nonmodifying], 25.2 [lib.alg.modifying.operations]  Status: Dup  Submitter: Peter Dimov  Date: 20 Sep 2004

c++std-lib-14262

[lib.alg.find] requires T to be EqualityComparable:

template <class InputIterator, class T>
   InputIterator find(InputIterator first, InputIterator last,
                      const T& value);

However the condition being tested, as specified in the Effects clause, is actually *i == value, where i is an InputIterator.

The two clauses are in agreement only if the type of *i is T, but this isn't necessarily the case. *i may have a heterogeneous comparison operator that takes a T, or a T may be convertible to the type of *i.

Further discussion (c++std-lib-14264): this problem affects a number of algorithsm in clause 25, not just find. We should try to resolve this problem everywhere it appears.

Proposed resolution:

[lib.alg.find]:

Remove [lib.alg.find]/1.

[lib.alg.count]:

Remove [lib.alg.count]/1.

[lib.alg.search]:

Remove "Type T is EqualityComparable (20.1.1), " from [lib.alg.search]/4.

[lib.alg.replace]:

Remove [lib.alg.replace]/1. Replace [lb.alg.replace]/2 with:

For every iterator i in the range [first, last) for which *i == value or pred(*i) holds perform *i = new_value.

Remove the first sentence of /4. Replace the beginning of /5 with:

For every iterator i in the range [result, result + (last - first)), assign to *i either...

(Note the defect here, current text says assign to i, not *i).

[lib.alg.fill]:

Remove "Type T is Assignable (23.1), " from /1. Replace /2 with:

For every iterator i in the range [first, last) or [first, first + n), perform *i = value.

[lib.alg.remove]:

Remove /1. Remove the first sentence of /6.

Rationale:

Duplicate of (a subset of) issue 283.


486. min/max CopyConstructible requirement is too strict

Section: 25.3.7 [lib.alg.min.max]  Status: Dup  Submitter: Dave Abrahams  Date: 13 Oct 2004

A straightforward implementation of these algorithms does not need to copy T.

Proposed resolution:

drop the the words "and CopyConstructible" from paragraphs 1 and 4

Rationale:

Dup of 281.


487. Allocator::construct is too limiting

Section: 20.1.5 [lib.allocator.requirements]  Status: NAD  Submitter: Dhruv Matani  Date: 17 Oct 2004

The standard's version of allocator::construct(pointer, const_reference) severely limits what you can construct using this function. Say you can construct a socket from a file descriptor. Now, using this syntax, I first have to manually construct a socket from the fd, and then pass the constructed socket to the construct() function so it will just to an uninitialized copy of the socket I manually constructed. Now it may not always be possible to copy construct a socket eh! So, I feel that the changes should go in the allocator::construct(), making it:

    template<typename T>
    struct allocator{
      template<typename T1>
      void construct(pointer T1 const& rt1);
    };

Now, the ctor of the class T which matches the one that takes a T1 can be called! Doesn't that sound great?

Proposed resolution:

Rationale:

NAD. STL uses copying all the time, and making it possible for allocators to construct noncopyable objects is useless in the absence of corresponding container changes. We might consider this as part of a larger redesign of STL.


489. std::remove / std::remove_if wrongly specified

Section: 25.2.7 [lib.alg.remove]  Status: NAD  Submitter: Thomas Mang  Date: 12 Dec 2004

In Section 25.2.7 [lib.alg.remove], paragraphs 1 to 5 describe the behavior of the mutating sequence operations std::remove and std::remove_if. However, the wording does not reflect the intended behavior [Note: See definition of intended behavior below] of these algorithms, as it is known to the C++ community [1].

1) Analysis of current wording:

25.2.7 [lib.alg.remove], paragraph 2:

Current wording says: "Effects: Eliminates all the elements referred to by iterator i in the range [first, last) for which the following corresponding conditions hold: *i == value, pred(*i) != false."

This sentences expresses specifically that all elements denoted by the (original) range [first, last) for which the corresponding condition hold will be eliminated. Since there is no formal definition of the term "eliminate" provided, the meaning of "eliminate" in everyday language implies that as postcondition, no element in the range denoted by [first, last) will hold the corresponding condition on reiteration over the range [first, last).

However, this is neither the intent [Note: See definition of intended behavior below] nor a general possible approach. It can be easily proven that if all elements of the original range[first, last) will hold the condition, it is not possible to substitute them by an element for which the condition will not hold.

25.2.7 [lib.alg.remove], paragraph 3:

Current wording says: "Returns: The end of the resulting range."

The resulting range is not specified. In combination with 25.2.7 [lib.alg.remove], paragraph 2, the only reasonable interpretation of this so-called resulting range is the range [first,last) - thus returning always the ForwardIterator 'last' parameter.

25.2.7 [lib.alg.remove], paragraph 4:

Current wording says: "Notes: Stable: the relative order of the elements that are not removed is the same as their relative order in the original range"

This sentences makes use of the term "removed", which is neither specified, nor used in a previous paragraph (which uses the term "eliminate"), nor unamgiuously separated from the name of the algorithm.

2) Description of intended behavior:

For the rest of this Defect Report, it is assumed that the intended behavior was that all elements of the range [first, last) which do not hold the condition *i == value (std::remove) or pred(*i) != false (std::remove_if)], call them s-elements [Note: s...stay], will be placed into a contiguous subrange of [first, last), denoted by the iterators [first, return value). The number of elements in the resulting range [first, return value) shall be equal to the number of s-elements in the original range [first, last). The relative order of the elements in the resulting subrange[first, return value) shall be the same as the relative order of the corresponding elements in the original range. It is undefined whether any elements in the resulting subrange [return value, last) will hold the corresponding condition, or not.

All implementations known to the author of this Defect Report comply with this intent. Since the intent of the behavior (contrary to the current wording) is also described in various utility references serving the C++ community [1], it is not expected that fixing the paragraphs will influence current code - unless the code relies on the behavior as it is described by current wording and the implementation indeed reflects the current wording, and not the intent.

3) Proposed fixes:

Change 25.2.7 [lib.alg.remove], paragraph 2 to:

"Effect: Places all the elements referred to by iterator i in the range [first, last) for which the following corresponding conditions hold : !(*i == value), pred(*i) == false into the subrange [first, k) of the original range, where k shall denote a value of type ForwardIterator. It is undefined whether any elements in the resulting subrange [k, last) will hold the corresponding condition, or not."

Comments to the new wording:

a) "Places" has no special meaning, and the everyday language meaning should fit. b) The corresponding conditions were negated compared to the current wording, becaue the new wording requires it. c) The wording "of the original range" might be redundant, since any subrange starting at 'first' and containing no more elements than the original range is implicitly a subrange of the original range [first, last). d) The iterator k was introduced instead of "return value" in order to avoid a cyclic dependency on 25.2.7/3. The wording ", where k shall denote a value of type ForwardIterator" might be redundant, because it follows implicitly by 25.2.7/3. e) "Places" does, in the author's opinion, explicitly forbid duplicating any element holding the corresponding condition in the original range [first, last) within the resulting range [first, k). If there is doubt this term might be not unambiguous regarding this, it is suggested that k is specified more closely by the following wording: "k shall denote a value of type ForwardIterator [Note: see d)] so that k - first is equal to the number of elements in the original range [first, last) for which the corresponding condition did hold". This could also be expressed as a separate paragraph "Postcondition:" f) The senctence "It is undefined whether any elements in the resulting subrange [k, last) will hold the corresponding condition, or not." was added consciously so the term "Places" does not imply if the original range [first, last) contains n elements holding the corresponding condition, the identical range[first, last) will also contain exactly n elements holding the corresponding condition after application of the algorithm.

Change 25.2.7 [lib.alg.remove], paragraph 3 to: "Returns: The iterator k."

Change 25.2.7 [lib.alg.remove], paragraph 4 to: "Notes: Stable: the relative order of the elements that are placed into the subrange [first, return value) shall be the same as their relative order was in the original range [first, last) prior to application of the algorithm."

Comments to the new wording:

a) the wording "was ... prior to application of the algorithm" is used to explicitly distinguish the original range not only by means of iterators, but also by a 'chronological' factor from the resulting range [first, return value). It might be redundant.

[1]: The wording of these references is not always unambiguous, and provided examples partially contradict verbal description of the algorithms, because the verbal description resembles the problematic wording of ISO/IEC 14882:2003.

Proposed resolution:

Rationale:

The LWG believes that the standard is sufficiently clear, and that there is no evidence of any real-world confusion about this point.


490. std::unique wrongly specified

Section: 25.2.8 [lib.alg.unique]  Status: NAD  Submitter: Thomas Mang  Date: 12 Dec 2004

In Section 25.2.8 [lib.alg.unique], paragraphs 1 to 3 describe the behavior of the mutating sequence operation std::unique. However, the wording does not reflect the intended behavior [Note: See definition of intended behavior below] of these algorithms, as it is known to the C++ community [1].

1) Analysis of current wording:

25.2.8 [lib.alg.unique], paragraph 1:

Current wording says: "Effects: Eliminates all but the first element from every consecutive group of equal elements referred to by the iterator i in the range [first, last) for which the following corresponding conditions hold: *i == *(i - 1) or pred(*i, *(i -1)) != false"

This sentences expresses specifically that all elements denoted by the (original) range [first, last) which are not but the first element from a consecutive group of equal elements (where equality is defined as *i == *(i - 1) or pred(*i, *(i - 1)) ! = false) [Note: See DR 202], call them r-elements [Note: r...remove], will be eliminated. Since there is no formal definition of the term "eliminate" provided, it is undefined how this "elimination" takes place. But the meaning of "eliminate" in everyday language seems to disallow explicitly that after application of the algorithm, any r-element will remain at any position of the range [first, last) [2].

Another defect in the current wording concerns the iterators used to compare two elements for equality: The current wording contains the expression "(i - 1)", which is not covered by 25/9 [Note: See DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions].

25.2.8 [lib.alg.unique], paragraph 2:

Current wording says: "Returns: The end of the resulting range."

The resulting range is not specified. In combination with 25.2.8 [lib.alg.unique], paragraph 1, one reasonable interpretation (in the author's opinion even the only possible interpretation) of this so-called resulting range is the range [first, last) - thus returning always the ForwardIterator 'last' parameter.

2) Description of intended behavior:

For the rest of this Defect Report, it is assumed that the intended behavior was that all elements denoted by the original range [first, last) which are the first element from a consecutive group of elements for which the corresponding conditions: *(i-1) == *i (for the version of unique without a predicate argument) or pred(*(i-1), *i) ! = false (for the version of unique with a predicate argument) [Note: If such a group of elements consists of only a single element, this is also considered the first element] [Note: See resolutions of DR 202], call them s-elements [Note: s...stay], will be placed into a contiguous subrange of [first, last), denoted by the iterators [first, return value). The number of elements in the resulting range [first, return value) shall be equal to the number of s-elements in the original range [first, last). Invalid iterator arithmetic expressions are expected to be resolved as proposed in DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions. It is also assumed by the author that the relative order of the elements in the resulting subrange [first, return value) shall be the same as the relative order of the corresponding elements (the s-elements) in the original range [Note: If this was not intended behavior, the additional proposed paragraph about stable order will certainly become obsolete]. Furthermore, the resolutions of DR 202 are partially considered.

All implementations known to the author of this Defect Report comply with this intent [Note: Except possible effects of DR 202]. Since this intent of the behavior (contrary to the current wording) is also described in various utility references serving the C++ community [1], it is not expected that fixing the paragraphs will influence current code [Note: Except possible effects of DR 202] - unless the code relies on the behavior as it is described by current wording and the implementation indeed reflects the current wording, and not the intent.

3) Proposed fixes:

Change 25.2.8 [lib.alg.unique], paragraph 1 to:

"Effect: Places the first element from every consecutive group of elements, referred to by the iterator i in the range [first, last), for which the following conditions hold: *(i-1) == *i (for the version of unique without a predicate argument) or pred(*(i -1), *i) != false (for the version of unique with a predicate argument), into the subrange [first, k) of the original range, where k shall denote a value of type ForwardIterator."

Comments to the new wording:

a) The new wording was influenced by the resolutions of DR 202. If DR 202 is resolved in another way, the proposed wording need also additional review. b) "Places" has no special meaning, and the everyday language meaning should fit. c) The expression "(i - 1)" was left, but is expected that DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions will take this into account. d) The wording "(for the version of unique without a predicate argument)" and "(for the version of unique with a predicate argument)" was added consciously for clarity and is in resemblence with current 23.2.2.4 [lib.list.ops], paragraph 19. It might be considered redundant. e) The wording "of the original range" might be redundant, since any subrange starting at first and containing no more elements than the original range is implicitly a subrange of the original range [first, last). f) The iterator k was introduced instead of "return value" in order to avoid a cyclic dependency on 25.2.8 [lib.alg.unique], paragraph 2. The wording ", where k shall denote a value of type ForwardIterator" might be redundant, because it follows implicitly by 25.2.8 [lib.alg.unique], paragraph 2. g) "Places" does, in the author's opinion, explicitly forbid duplicating any s-element in the original range [first, last) within the resulting range [first, k). If there is doubt this term might be not unambiguous regarding this, it is suggested that k is specified more closely by the following wording: "k shall denote a value of type ForwardIterator [Note: See f)] so that k - first is equal to the number of elements in the original range [first, last) being the first element from every consecutive group of elements for which the corresponding condition did hold". This could also be expressed as a separate paragraph "Postcondition:". h) If it is considered that the wording is unclear whether it declares the element of a group which consists of only a single element implicitly to be the first element of this group [Note: Such an interpretation could eventually arise especially in case last - first == 1] , the following additional sentence is proposed: "If such a group of elements consists of only a single element, this element is also considered the first element."

Change 25.2.8 [lib.alg.unique], paragraph 2 to: "Returns: The iterator k."

Add a separate paragraph "Notes:" as 25.2.8 [lib.alg.unique], paragraph 2a or 3a, or a separate paragraph "Postcondition:" before 25.2.8 [lib.alg.unique], paragraph 2 (wording inside {} shall be eliminated if the preceding expressions are used, or the preceding expressions shall be eliminated if wording inside {} is used):

"Notes:{Postcondition:} Stable: the relative order of the elements that are placed into the subrange [first, return value {k}) shall be the same as their relative order was in the original range [first, last) prior to application of the algorithm."

Comments to the new wording:

a) It is assumed by the author that the algorithm was intended to be stable. In case this was not the intent, this paragraph becomes certainly obsolete. b) The wording "was ... prior to application of the algorithm" is used to explicitly distinguish the original range not only by means of iterators, but also by a 'chronological' factor from the resulting range [first, return value). It might be redundant.

25.2.8 [lib.alg.unique], paragraph 3:

See DR 239.

4) References to other DRs:

See DR 202, but which does not address any of the problems described in this Defect Report [Note: This DR is supposed to complement DR 202]. See DR 239. See DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions.

[1]: The wording of these references is not always unambiguous, and provided examples partially contradict verbal description of the algorithms, because the verbal description resembles the problematic wording of ISO/IEC 14882:2003.

[2]: Illustration of conforming implementations according to current wording:

One way the author of this DR considers how this "elimination" could be achieved by a conforming implementation according to current wording is by substituting each r-element by _any_ s-element [Note: s...stay; any non-r-element], since all r-elements are "eliminated".

In case of a sequence consisting of elements being all 'equal' [Note: See DR 202], substituting each r-element by the single s-element is the only possible solution according to current wording.

Proposed resolution:

Rationale:

The LWG believes the standard is sufficiently clear. No implementers get it wrong, and changing it wouldn't cause any code to change, so there is no real-world harm here.


491. std::list<>::unique incorrectly specified

Section: 23.2.2.4 [lib.list.ops]  Status: NAD  Submitter: Thomas Mang  Date: 12 Dec 2004

In Section 23.2.2.4 [lib.list.ops], paragraphs 19 to 21 describe the behavior of the std::list<T, Allocator>::unique operation. However, the current wording is defective for various reasons.

1) Analysis of current wording:

23.2.2.4 [lib.list.ops], paragraph 19:

Current wording says: "Effects: Eliminates all but the first element from every consecutive group of equal elements referred to by the iterator i in the range [first + 1, last) for which *i == *(i - 1) (for the version of unique with no argument) or pred(*i, *(i -1)) (for the version of unique with a predicate argument) holds."

This sentences makes use of the undefined term "Eliminates". Although it is, to a certain degree, reasonable to consider the term "eliminate" synonymous with "erase", using "Erase" in the first place, as the wording of 23.2.2.4 [lib.list.ops], paragraph 15 does, would be clearer.

The range of the elements referred to by iterator i is "[first + 1, last)". However, neither "first" nor "last" is defined.

The sentence makes three times use of iterator arithmetic expressions ( "first + 1", "*i == *(i - 1)", "pred(*i, *(i -1))" ) which is not defined for bidirectional iterator [see DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions].

The same problems as pointed out in DR 202 (equivalence relation / order of arguments for pred()) apply to this paragraph.

23.2.2.4 [lib.list.ops], paragraph 20:

Current wording says: "Throws: Nothing unless an exception in thrown by *i == *(i-1) or pred(*i, *(i - 1))"

The sentence makes two times use of invalid iterator arithmetic expressions ( "*i == *(i - 1)", "pred(*i, *(i -1))" ).

[Note: Minor typos: "in" / missing dot at end of sentence.]

23.2.2.4 [lib.list.ops], paragraph 21:

Current wording says: "Complexity: If the range (last - first) is not empty, exactly (last - first) - 1 applications of the corresponding predicate, otherwise no application of the predicate.

See DR 315 regarding "(last - first)" not yielding a range.

Invalid iterator arithmetic expression "(last - first) - 1" left .

2) Description of intended behavior:

For the rest of this Defect Report, it is assumed that "eliminate" is supposed to be synonymous to "erase", that "first" is equivalent to an iterator obtained by a call to begin(), "last" is equivalent to an iterator obtained by a call to end(), and that all invalid iterator arithmetic expressions are resolved as described in DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions.

Furthermore, the resolutions of DR 202 are considered regarding equivalence relation and order of arguments for a call to pred.

All implementations known to the author of this Defect Report comply with these assumptions, apart from the impact of the alternative resolution of DR 202. Except for the changes implied by the resolutions of DR 202, no impact on current code is expected.

3) Proposed fixes:

Change 23.2.2.4 [lib.list.ops], paragraph 19 to:

"Effect: Erases all but the first element from every consecutive group of elements, referred to by the iterator i in the range [begin(), end()), for which the following conditions hold: *(i-1) == *i (for the version of unique with no argument) or pred(*(i-1), *i) != false (for the version of unique with a predicate argument)."

Comments to the new wording:

a) The new wording was influenced by DR 202 and the resolutions presented there. If DR 202 is resolved in another way, the proposed wording need also additional review. b) "Erases" refers in the author's opinion unambiguously to the member function "erase". In case there is doubt this might not be unamgibuous, a direct reference to the member function "erase" is suggested [Note: This would also imply a change of 23.2.2.4 [lib.list.ops], paragraph 15.]. c) The expression "(i - 1)" was left, but is expected that DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions will take this into account. d) The wording "(for the version of unique with no argument)" and "(for the version of unique with a predicate argument)" was kept consciously for clarity. e) "begin()" substitutes "first", and "end()" substitutes "last". The range need adjustment from "[first + 1, last)" to "[begin(), end())" to ensure a valid range in case of an empty list. f) If it is considered that the wording is unclear whether it declares the element of a group which consists of only a single element implicitly to be the first element of this group [Note: Such an interpretation could eventually arise especially in case size() == 1] , the following additional sentence is proposed: "If such a group of elements consists of only a single element, this element is also considered the first element."

Change 23.2.2.4 [lib.list.ops], paragraph 20 to:

"Throws: Nothing unless an exception is thrown by *(i-1) == *i or pred(*(i-1), *i)."

Comments to the new wording:

a) The wording regarding the conditions is identical to proposed 23.2.2.4 [lib.list.ops], paragraph 19. If 23.2.2.4 [lib.list.ops], paragraph 19 is resolved in another way, the proposed wording need also additional review. b) The expression "(i - 1)" was left, but is expected that DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions will take this into account. c) Typos fixed.

Change 23.2.2.4 [lib.list.ops], paragraph 21 to:

"Complexity: If empty() == false, exactly size() - 1 applications of the corresponding predicate, otherwise no applications of the corresponding predicate."

Comments to the new wording:

a) The new wording is supposed to also replace the proposed resolution of DR 315, which suffers from the problem of undefined "first" / "last".

5) References to other DRs:

See DR 202. See DR 239. See DR 315. See DR submitted by Thomas Mang regarding invalid iterator arithmetic expressions.

Proposed resolution:

Rationale:

"All implementations known to the author of this Defect Report comply with these assumption", and "no impact on current code is expected", i.e. there is no evidence of real-world confusion or harm.


493. Undefined Expression in Input Iterator Note Title

Section: 24.1.1 [lib.input.iterators]  Status: NAD  Submitter: Chris Jefferson  Date: 13 Dec 2004

1) In 24.1.1/3, the following text is currently present.

"Note: For input iterators, a==b does not imply ++a=++b (Equality does not guarantee the substitution property or referential transparency)."

However, when in Table 72, part of the definition of ++r is given as:

"pre: r is dereferenceable. post: any copies of the previous value of r are no longer required either to be dereferenceable ..."

While a==b does not imply that b is a copy of a, this statement should perhaps still be made more clear.

2) There are no changes to intended behaviour

3) This Note should be altered to say "Note: For input iterators a==b, when its behaviour is defined ++a==++b may still be false (Equality does not guarantee the substitution property or referential transparency).

Proposed resolution:

Rationale:

This is descriptive text, not normative, and the meaning is clear.


494. Wrong runtime complexity for associative container's insert and delete

Section: 23.1.2 [lib.associative.reqmts]  Status: NAD  Submitter: Hans B os  Date: 19 Dec 2004

According to [lib.associative.reqmts] table 69, the runtime comlexity of insert(p, t) and erase(q) can be done in amortized constant time.

It was my understanding that an associative container could be implemented as a balanced binary tree.

For inser(p, t), you 'll have to iterate to p's next node to see if t can be placed next to p. Furthermore, the insertion usually takes place at leaf nodes. An insert next to the root node will be done at the left of the root next node

So when p is the root node you 'll have to iterate from the root to its next node, which takes O(log(size)) time in a balanced tree.

If you insert all values with insert(root, t) (where root is the root of the tree before insertion) then each insert takes O(log(size)) time. The amortized complexity per insertion will be O(log(size)) also.

For erase(q), the normal algorithm for deleting a node that has no empty left or right subtree, is to iterate to the next (or previous), which is a leaf node. Then exchange the node with the next and delete the leaf node. Furthermore according to DR 130, erase should return the next node of the node erased. Thus erasing the root node, requires iterating to the next node.

Now if you empty a map by deleting the root node until the map is empty, each operation will take O(log(size)), and the amortized complexity is still O(log(size)).

The operations can be done in amortized constant time if iterating to the next node can be done in (non amortized) constant time. This can be done by putting all nodes in a double linked list. This requires two extra links per node. To me this is a bit overkill since you can already efficiently insert or erase ranges with erase(first, last) and insert(first, last).

Proposed resolution:

Rationale:

Only "amortized constant" in special circumstances, and we believe that's implementable. That is: doing this N times will be O(N), not O(log N).

----- End of document -----